icm

[English](README.md) | [Français](README_fr.md) | [Español](README_es.md) | [Deutsch](README_de.md) | [Italiano](README_it.md) | [Português](README_pt.md) | [Nederlands](README_nl.md) | [Polski](README_pl.md) | [Русский](README_ru.md) | [日本語](README_ja.md) | [中文](README_zh.md) | [العربية](README_ar.md) | [한국어](README_ko.md) <p align="center"> <img src="assets/banner.png" alt="ICM — Infinite Context Memory" width="600"> </p> <h1 align="center">ICM</h1> <p align="center"> Permanent memory for AI agents. Single binary, zero dependencies, MCP native. </p> <p align="center"> <a href="https://github.com/rtk-ai/icm/actions/workflows/ci.yml"><img src="https://github.com/rtk-ai/icm/actions/workflows/ci.yml/badge.svg" alt="CI"></a> <a href="https://github.com/rtk-ai/icm/releases/latest"><img src="https://img.shields.io/github/v/release/rtk-ai/icm?color=purple" alt="Release"></a> <a href="LICENSE"><img src="https://img.shields.io/badge/License-Apache--2.0-blue.svg" alt="Apache-2.0"></a> </p> --- > ⚠️ **Project status: experimental** > > ICM is pre-1.0 and under active development. Breaking changes can land > in any minor release, and hooks/MCP configuration formats may shift. > > That said, I (the maintainer) use ICM every day as my primary AI > coding memory layer — it's experimental in API stability, not in > day-to-day usefulness. > > My focus is currently on [rtk](https://github.com/rtk-ai/rtk); ICM > updates are merged on a best-effort cadence. Issues and pull requests > are welcome but may take longer to review than usual. > > ICM is Apache-2.0 licensed and ships **as-is, without warranty of any > kind** (see [LICENSE](LICENSE)). Before any destructive operation, run > the read-only equivalent first (`icm uninstall --dry-run`, > `icm uninstall --check`). --- ICM gives your AI agent a real memory — not a note-taking tool, not a context manager, a **memory**. ``` ICM (Infinite Context Memory) ┌──────────────────────┬──────────────────────────┐ │ MEMORIES (Topics) │ MEMOIRS (Knowledge) │ │ │ │ │ Episodic, temporal │ Permanent, structured │ │ │ │ │ ┌───┐ ┌───┐ ┌───┐ │ ┌───┐ │ │ │ m │ │ m │ │ m │ │ │ C │──depends_on──┐ │ │ └─┬─┘ └─┬─┘ └─┬─┘ │ └───┘ │ │ │ │decay│ │ │ │ refines │ │ │ ▼ ▼ ▼ │ ┌─▼─┐ ┌─▼─┐│ │ weight decreases │ │ C │──part_of──>│ C ││ │ over time unless │ └───┘ └───┘│ │ accessed/critical │ Concepts + Relations │ ├──────────────────────┴──────────────────────────┤ │ SQLite + FTS5 + sqlite-vec │ │ Hybrid search: BM25 (30%) + cosine (70%) │ └─────────────────────────────────────────────────┘ ``` **Two memory models:** - **Memories** — store/recall with temporal decay by importance. Critical memories never fade, low-importance ones decay naturally. Filter by topic or keyword. - **Memoirs** — permanent knowledge graphs. Concepts linked by typed relations (`depends_on`, `contradicts`, `superseded_by`, ...). Filter by label. - **Feedback** — record corrections when AI predictions are wrong. Search past mistakes before making new predictions. Closed-loop learning. ## One memory, every AI tool — no re-explaining The point of ICM: **you stop re-explaining context every time you switch AI tools**. Tell Claude Code about your project's auth strategy on Monday — Tuesday's Gemini session already knows it. Hit a Postgres indexing gotcha in a Codex run — next week's Cursor session finds the fix in `errors-resolved`. This works because every AI tool you configure with `icm init` reads and writes the **same SQLite database** at the OS-standard data location (e.g. `~/.local/share/icm/memories.db` on Linux, `~/Library/Application Support/icm/memories.db` on macOS, `%APPDATA%\icm\icm\data\memories.db` on Windows): - A `icm store -t decisions-myapp -c "..."` from Claude Code is immediately visible to Codex, Gemini, Cursor, Roo, Amp, Aider, ... - `icm recall "query"` from any tool searches the same corpus. - Topics (`decisions-myapp`, `preferences`, `errors-resolved`, ...) are global — there is no per-tool partition. The [multi-agent benchmark](#multi-agent-unified-memory) below confirms it end-to-end: facts seeded through ICM are recalled with 100% accuracy by Claude Code, Gemini CLI, Copilot CLI, Cursor Agent, and Aider — **98% cross-agent efficiency** on the standard test. If you want isolation (per-project, per-tool, etc.) pass `--db <path>` or set `ICM_DB_PATH`; each path is an independent corpus. ## Install ```bash # Homebrew (macOS / Linux) brew tap rtk-ai/tap && brew install icm # Quick install (macOS / Linux) — verifies SHA256 against the release checksums curl -fsSL https://raw.githubusercontent.com/rtk-ai/icm/main/install.sh | sh # Quick install (Windows PowerShell) irm https://raw.githubusercontent.com/rtk-ai/icm/main/install.ps1 | iex # From source cargo install --path crates/icm-cli ``` Re-run the install command to upgrade to the latest release. To pin a version, pass `--version icm-vX.Y.Z` (sh: `sh -s -- --version …`). ## Setup ```bash # Auto-detect and configure all supported tools icm init ``` Configures **18 tools** in one command ([full integration guide](docs/integrations.md)): | Tool | MCP | Hooks | CLI | Skills | |------|:---:|:-----:|:---:|:------:| | Claude Code | `~/.claude.json` | 5 hooks | `CLAUDE.md` | `/recall` `/remember` | | Claude Desktop | JSON | — | — | — | | Gemini CLI | `~/.gemini/settings.json` | 5 hooks | `GEMINI.md` | — | | Codex CLI | `~/.codex/config.toml` | 3 hooks (PostToolUse opt-in, see #288) | `AGENTS.md` | — | | Copilot CLI | `~/.copilot/mcp-config.json` | 4 hooks | `.github/copilot-instructions.md` | — | | Cursor | `~/.cursor/mcp.json` | — | — | `.mdc` rule | | Windsurf | JSON | — | `.windsurfrules` | — | | VS Code | `~/Library/.../Code/User/mcp.json` | — | — | — | | Amp | JSON | — | — | `/icm-recall` `/icm-remember` | | Amazon Q | JSON | — | — | — | | Cline | VS Code globalStorage | — | — | — | | Roo Code | VS Code globalStorage | — | — | `.md` rule | | Kilo Code | VS Code globalStorage | — | — | — | | Zed | `~/.zed/settings.json` | — | — | — | | OpenCode | JSON | TS plugin | — | — | | Continue.dev | `~/.continue/config.yaml` | — | — | — | | Aider | — | — | `.aider.conventions.md` | — | | Pi | — | TS ext (TBD) | `~/.pi/agent/AGENTS.md` | `/icm-recall` `/icm-remember` | Or manually: ```bash # Claude Code claude mcp add icm -- icm serve # Compact mode (shorter responses, saves tokens) claude mcp add icm -- icm serve --compact # Any MCP client: command = "icm", args = ["serve"] ``` ### Skills / rules ```bash icm init --mode skill ``` Installs slash commands and rules for Claude Code (`/recall`, `/remember`), Cursor (`.mdc` rule), Roo Code (`.md` rule), and Amp (`/icm-recall`, `/icm-remember`). ### CLI instructions ```bash icm init --mode cli ``` Injects ICM instructions into each tool's instruction file: | Tool | File | |------|------| | Claude Code | `CLAUDE.md` | | GitHub Copilot | `.github/copilot-instructions.md` | | Windsurf | `.windsurfrules` | | OpenAI Codex | `AGENTS.md` | | Gemini | `~/.gemini/GEMINI.md` | ### Hooks (5 tools) ```bash icm init --mode hook ``` Installs auto-extraction and auto-recall hooks for all supported tools: | Tool | SessionStart | PreTool | PostTool | Compact | PromptRecall | Config | |------|:-----------:|:-------:|:--------:|:-------:|:------------:|--------| | Claude Code | `icm hook start` | `icm hook pre` | `icm hook post` | `icm hook compact` | `icm hook prompt` | `~/.claude/settings.json` | | Gemini CLI | `icm hook start` | `icm hook pre` | `icm hook post` | `icm hook compact` | `icm hook prompt` | `~/.gemini/settings.json` | | Codex CLI | `icm hook start` | `icm hook pre` | `icm hook post`¹ | — | `icm hook prompt` | `~/.codex/hooks.json` | | Copilot CLI | `icm hook start` | `icm hook pre` | `icm hook post` | — | `icm hook prompt` | `.github/hooks/icm.json` | | OpenCode | session start | — | tool extract | compaction | — | `~/.config/opencode/plugins/icm.ts` | **What each hook does:** | Hook | What it does | |------|-------------| | `icm hook start` | Inject a wake-up pack of critical/high memories at session start (~500 tokens) | | `icm hook pre` | Auto-allow `icm` CLI commands (no permission prompt) | | `icm hook post` | Extract facts from tool output every N calls (auto-extraction) | | `icm hook compact` | Extract memories from transcript before context compression | | `icm hook prompt` | Inject recalled context at the start of each user prompt | ¹ **Codex CLI PostToolUse is off by default.** Codex fires PostToolUse on every shell command — a session generates ~14k events / 24h, which floods the store with tool-output bloat (issue #288). Opt in with `icm init --with-codex-post-hook` if you want it; tune `[extraction]` first (`extract_every`, `min_score`, `store_raw = false`). MCP + `AGENTS.md` alone still let Codex save via the `icm_memory_store` tool. ## CLI vs MCP ICM can be used via CLI (`icm` commands) or MCP server (`icm serve`). Both access the same database. | | CLI | MCP | |---|-----|-----| | **Latency** | ~30ms (direct binary) | ~50ms (JSON-RPC stdio) | | **Token cost** | 0 (hook-based, invisible) | ~20-50 tokens/call (tool schema) | | **Setup** | `icm init --mode hook` | `icm init --mode mcp` | | **Works with** | Claude Code, Gemini, Codex, Copilot, OpenCode (via hooks) | All 17 MCP-compatible tools | | **Auto-extraction** | Yes (hooks trigger `icm extract`) | Yes (MCP tools call store) | | **Best for** | Power users, token savings | Universal compatibility | ## HTTP API (warm model) ```bash # Persistent local server — embedding model loads once, stays warm. icm serve --http 127.0.0.1:11435 --db ~/.local/share/icm/memories.db & curl -s -X POST 127.0.0.1:11435/store \ -H 'content-type: application/json' \ -d '{"topic":"t","content":"hello world","keywords":"x"}' # TOON by default (lowest token cost on LLM-side reads). curl -s -X POST 127.0.0.1:11435/recall \ -H 'content-type: application/json' \ -d '{"query":"hello","topic":"t","limit":5}' # JSON variant: ?format=json or Accept: application/json curl -s -X POST '127.0.0.1:11435/recall?format=json' \ -H 'content-type: application/json' \ -d '{"query":"hello","topic":"t"}' ``` Endpoints: `POST /store`, `POST /recall`, `POST /consolidate`, `GET /stats`, `GET /topics`, `GET /health`. Optional `--token <T>` enables `Authorization: Bearer <T>` on every request (health stays open as a liveness probe). Bound to whatever address you pass; `127.0.0.1:<port>` keeps the server localhost-only. Saves ~9 s per call vs one-shot CLI (model reload) — any scripting language can hit semantic recall with plain `curl`. Requires the `http-api` feature (enabled by default). Issue [#290](https://github.com/rtk-ai/icm/issues/290). ## Dashboard ```bash icm dashboard # or: icm tui ``` Interactive TUI with 5 tabs: Overview, Topics, Memories, Health, Memoirs. Keyboard navigation (vim-style: j/k, g/G, Tab, 1-5), live search (/), auto-refresh. Requires the `tui` feature (enabled by default). Build without: `cargo install --path crates/icm-cli --no-default-features --features embeddings`. ## CLI ### Memories (episodic, with decay) ```bash # Store icm store -t "my-project" -c "Use PostgreSQL for the main DB" -i high -k "db,postgres" # Recall icm recall "database choice" icm recall "auth setup" --topic "my-project" --limit 10 icm recall "architecture" --keyword "postgres" # Manage icm forget <memory-id> icm consolidate --topic "my-project" icm topics icm stats # Extract facts from text (rule-based, zero LLM cost) echo "The parser uses Pratt algorithm" | icm extract -p my-project ``` ### Memoirs (permanent knowledge graphs) ```bash # Create a memoir icm memoir create -n "system-architecture" -d "System design decisions" # Add concepts with labels icm memoir add-concept -m "system-architecture" -n "auth-service" \ -d "Handles JWT tokens and OAuth2 flows" -l "domain:auth,type:service" # Link concepts icm memoir link -m "system-architecture" --from "api-gateway" --to "auth-service" -r depends-on # Search with label filter icm memoir search -m "system-architecture" "authentication" icm memoir search -m "system-architecture" "service" --label "domain:auth" # Inspect neighborhood icm memoir inspect -m "system-architecture" "auth-service" -D 2 # Export graph (formats: json, dot, ascii, ai) icm memoir export -m "system-architecture" -f ascii # Box-drawing with confidence bars icm memoir export -m "system-architecture" -f dot # Graphviz DOT (color = confidence level) icm memoir export -m "system-architecture" -f ai # Markdown optimized for LLM context icm memoir export -m "system-architecture" -f json # Structured JSON with all metadata # Generate SVG visualization icm memoir export -m "system-architecture" -f dot | dot -Tsvg > graph.svg ``` ### Transcripts (verbatim session replay) Store every message exchanged with an agent as-is — no summarization, no extraction. Search later with FTS5 (BM25 + boolean + phrase + prefix). Useful for session replay, post-mortem review, compliance audit, training data. Complementary to curated memories. ```bash # 1. Start a session SID=$(icm transcript start-session --agent claude-code --project myapp) # 2. Record every turn verbatim icm transcript record -s "$SID" -r user -c "Pourquoi on avait choisi Postgres ?" icm transcript record -s "$SID" -r assistant -c "JSONB natif, BRIN pour les logs, auto-vacuum tuné." icm transcript record -s "$SID" -r tool -c '{"cmd":"psql -c ..."}' -t Bash --tokens 42 # 3. Replay, search, inspect icm transcript list-sessions --project myapp icm transcript show "$SID" --limit 200 icm transcript search "postgres JSONB" # BM25 ranked icm transcript search '"auto-vacuum"' # phrase match icm transcript search "postgres OR mysql" --session "$SID" # boolean, scoped icm transcript stats # 4. Delete a session (cascade deletes its messages) icm transcript forget "$SID" ``` Rust + SQLite + FTS5 — 0 Python, 0 ChromaDB, 0 external service. Writes are ~10× faster than ChromaDB-based verbatim stores; the whole transcript lives in the same SQLite file as your memories and memoirs. ## MCP Tools (31) ### Memory tools | Tool | Description | |------|-------------| | `icm_memory_store` | Store with auto-dedup (>85% similarity → update instead of duplicate) | | `icm_memory_recall` | Search by query, filter by topic / keyword / project | | `icm_memory_update` | Edit a memory in-place (content, importance, keywords) | | `icm_memory_forget` | Delete a memory by ID | | `icm_memory_forget_topic` | Delete all memories in a given topic | | `icm_memory_consolidate` | Merge all memories of a topic into one summary | | `icm_memory_extract_patterns` | Detect recurring patterns within a topic and surface them as concepts | | `icm_memory_list_topics` | List all topics with counts | | `icm_memory_stats` | Global memory statistics | | `icm_memory_health` | Per-topic hygiene audit (staleness, consolidation needs) | | `icm_memory_embed_all` | Backfill embeddings for vector search | ### Session tools | Tool | Description | |------|-------------| | `icm_wake_up` | Build a project-scoped wake-up pack (critical/high memories + preferences) for SessionStart-style context injection | | `icm_learn` | Scan a project directory and seed a Memoir knowledge graph from its code/docs | ### Memoir tools (knowledge graphs) | Tool | Description | |------|-------------| | `icm_memoir_create` | Create a new memoir (knowledge container) | | `icm_memoir_list` | List all memoirs | | `icm_memoir_show` | Show memoir details and all concepts | | `icm_memoir_add_concept` | Add a concept with labels | | `icm_memoir_refine` | Update a concept's definition | | `icm_memoir_search` | Full-text search, optionally filtered by label | | `icm_memoir_search_all` | Search across all memoirs | | `icm_memoir_link` | Create typed relation between concepts | | `icm_memoir_inspect` | Inspect concept and graph neighborhood (BFS) | | `icm_memoir_export` | Export graph (json, dot, ascii, ai) with confidence levels | ### Feedback tools (learning from mistakes) | Tool | Description | |------|-------------| | `icm_feedback_record` | Record a correction when an AI prediction was wrong | | `icm_feedback_search` | Search past corrections to inform future predictions | | `icm_feedback_stats` | Feedback statistics: total count, breakdown by topic, most applied | ### Transcript tools (verbatim session replay) | Tool | Description | |------|-------------| | `icm_transcript_start_session` | Create a session for verbatim message capture; returns `session_id` | | `icm_transcript_record` | Append a raw message (role, content, optional tool + tokens + metadata) | | `icm_transcript_search` | FTS5 search across messages (BM25, boolean, phrase, prefix) | | `icm_transcript_show` | Replay full message thread of a session, chronologically | | `icm_transcript_stats` | Sessions, messages, bytes, breakdown by role/agent/top-sessions | ### Relation types `part_of` · `depends_on` · `related_to` · `contradicts` · `refines` · `alternative_to` · `caused_by` · `instance_of` · `superseded_by` ## How it works ### Dual memory model **Episodic memory (Topics)** captures decisions, errors, preferences. Each memory has a weight that decays over time based on importance: | Importance | Decay | Prune | Behavior | |-----------|-------|-------|----------| | `critical` | none | never | Never forgotten, never pruned | | `high` | slow (0.5x rate) | never | Fades slowly, never auto-deleted | | `medium` | normal | yes | Standard decay, pruned when weight < threshold | | `low` | fast (2x rate) | yes | Quickly forgotten | Decay is **access-aware**: frequently recalled memories decay slower (`decay / (1 + access_count × 0.1)`). Applied automatically on recall (if >24h since last decay). **Memory hygiene** is built-in: - **Auto-dedup**: storing content >85% similar to an existing memory in the same topic updates it instead of creating a duplicate - **Consolidation hints**: when a topic exceeds 7 entries, `icm_memory_store` warns the caller to consolidate - **Health audit**: `icm_memory_health` reports per-topic entry count, average weight, stale entries, and consolidation needs - **No silent data loss**: critical and high-importance memories are never auto-pruned **Semantic memory (Memoirs)** captures structured knowledge as a graph. Concepts are permanent — they get refined, never decayed. Use `superseded_by` to mark obsolete facts instead of deleting them. ### Hybrid search With embeddings enabled, ICM uses hybrid search: - **FTS5 BM25** (30%) — full-text keyword matching - **Cosine similarity** (70%) — semantic vector search via sqlite-vec Default model: `intfloat/multilingual-e5-base` (768d, 100+ languages). Configurable in your [config file](#configuration): ```toml [embeddings] # enabled = false # Disable entirely (no model download) model = "intfloat/multilingual-e5-base" # 768d, multilingual (default) # model = "intfloat/multilingual-e5-small" # 384d, multilingual (lighter) # model = "intfloat/multilingual-e5-large" # 1024d, multilingual (best accuracy) # model = "Xenova/bge-small-en-v1.5" # 384d, English-only (fastest) # model = "jinaai/jina-embeddings-v2-base-code" # 768d, code-optimized ``` To skip the embedding model download entirely, use any of these: ```bash icm --no-embeddings serve # CLI flag ICM_NO_EMBEDDINGS=1 icm serve # Environment variable ``` Or set `enabled = false` in your config file. ICM will fall back to FTS5 keyword search (still works, just no semantic matching). Changing the model automatically re-creates the vector index (existing embeddings are cleared and can be regenerated with `icm_memory_embed_all`). ### Storage Single SQLite file. No external services, no network dependency. ``` ~/Library/Application Support/dev.icm.icm/memories.db # macOS ~/.local/share/dev.icm.icm/memories.db # Linux C:\Users\<user>\AppData\Local\icm\icm\data\memories.db # Windows ``` ### Configuration ```bash icm config # Show active config ``` Config file location (platform-specific, or `$ICM_CONFIG`): ``` ~/Library/Application Support/dev.icm.icm/config.toml # macOS ~/.config/icm/config.toml # Linux C:\Users\<user>\AppData\Roaming\icm\icm\config\config.toml # Windows ``` See [config/default.toml](config/default.toml) for all options. ## Multi-project & multi-agent ICM is built for the case where one user collaborates with many agents across many projects. Memories must stay relevant: a decision from project A should never leak into project B, and a `dev` agent should not be hydrated with what a `mentor` agent stored. ### Project isolation ICM scopes memories by **topic naming convention**, not by a separate column. The convention: ``` {kind}-{project} # e.g. decisions-icm, errors-resolved-icm, contexte-rtk-cloud preferences # global, always included identity # global, always included ``` `icm_wake_up { project: "icm" }` does **segment-aware** matching: `"icm"` matches `decisions-icm`, `errors-icm-core`, `contexte-icm` — but never `icmp-notes` (no false positives). Topics are split on `-`, `.`, `_`, `/`, `:`. Preference and identity topics are cross-project by design — user-level guidance is never stripped. The `UserPromptSubmit` hook (`icm hook prompt`) and the `SessionStart` hook (`icm hook start`) both derive the project from the `cwd` field in the hook JSON (`basename` of the working directory). Run each project from its own directory and isolation is automatic. ### Writing good memories `icm_memory_store` requires the agent to choose `topic` and `content` — there is no auto-classifier. Best practice: | Field | Guidance | |------|----------| | `topic` | `{kind}-{project}`. Kinds: `decisions`, `errors-resolved`, `contexte`, `preferences`. | | `content` | One fact per store. Dense English summary — `topic + content` is the embedding text. | | `raw_excerpt` | Verbatim only (code, exact error message, command output). | | `keywords` | 3–5 terms to boost BM25 retrieval. | | `importance` | `critical` for never-forget, `high` for project decisions, `medium` default, `low` for ephemeral. | ICM handles the rest: **dedup at 85% similarity**, **auto-link** between semantically close memories, **auto-consolidation** above 10 entries per topic, and **decay** weighted by access count. One fact per call beats batched dumps — the retriever ranks individually-stored facts higher. ### Multi-agent roles ICM does not yet have a first-class `role` column. Today, roles are emulated by topic suffixes plus per-agent working directories: ``` decisions-icm-dev # dev agent: code patterns, library choices, refactors decisions-icm-architect # architect: design, workflows, subtask decomposition decisions-icm-mentor # mentor / BA: business goals, non-technical context ``` Each agent runs in its own working directory (`~/projects/icm-dev/`, `~/projects/icm-architect/`, ...) so that `icm hook prompt` and `icm hook start` derive a different project segment from `cwd` and only recall the matching memories. Preferences remain global — user identity carries across all roles. Within a single agent, you can also narrow recall manually: ```jsonc // icm_memory_recall { "query": "auth flow", "topic": "decisions-icm-architect", "limit": 5 } ``` A first-class `role` field (with native filtering in wake-up and recall) is on the roadmap. Until then, the topic-suffix convention is the supported pattern. ## Auto-extraction ICM extracts memories automatically via three layers: ``` Layer 0: Pattern hooks Layer 1: PreCompact Layer 2: UserPromptSubmit (zero LLM cost) (zero LLM cost) (zero LLM cost) ┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐ │ PostToolUse hook │ │ PreCompact hook │ │ UserPromptSubmit │ │ │ │ │ │ │ │ • Bash errors │ │ Context about to │ │ User sends prompt │ │ • git commits │ │ be compressed → │ │ → icm recall │ │ • config changes │ │ extract memories │ │ → inject context │ │ • decisions │ │ from transcript │ │ │ │ • preferences │ │ before they're │ │ Agent starts with │ │ • learnings │ │ lost forever │ │ relevant memories │ │ • constraints │ │ │ │ already loaded │ │ │ │ Same patterns + │ │ │ │ Rule-based, no LLM│ │ --store-raw fallbk│ │ │ └──────────────────┘ └──────────────────┘ └──────────────────┘ ``` | Layer | Status | LLM cost | Hook command | Description | |-------|--------|----------|-------------|-------------| | Layer 0 | Implemented | 0 | `icm hook post` | Rule-based keyword extraction from tool output | | Layer 1 | Implemented | 0 | `icm hook compact` | Extract from transcript before context compression | | Layer 2 | Implemented | 0 | `icm hook prompt` | Inject recalled memories on each user prompt | All 3 layers are installed automatically by `icm init --mode hook`. ## Benchmarks ### Storage performance ``` ICM Benchmark (1000 memories, 384d embeddings) ────────────────────────────────────────────────────────── Store (no embeddings) 1000 ops 34.2 ms 34.2 µs/op Store (with embeddings) 1000 ops 51.6 ms 51.6 µs/op FTS5 search 100 ops 4.7 ms 46.6 µs/op Vector search (KNN) 100 ops 59.0 ms 590.0 µs/op Hybrid search 100 ops 95.1 ms 951.1 µs/op Decay (batch) 1 ops 5.8 ms 5.8 ms/op ────────────────────────────────────────────────────────── ``` Apple M1 Pro, in-memory SQLite, single-threaded. `icm bench --count 1000` ### Agent efficiency Multi-session workflow with a real Rust project (12 files, ~550 lines). Sessions 2+ show the biggest gains as ICM recalls instead of re-reading files. ``` ICM Agent Benchmark (10 sessions, model: haiku, 3 runs averaged) ══════════════════════════════════════════════════════════════════ Without ICM With ICM Delta Session 2 (recall) Turns 5.7 4.0 -29% Context (input) 99.9k 67.5k -32% Cost $0.0298 $0.0249 -17% Session 3 (recall) Turns 3.3 2.0 -40% Context (input) 74.7k 41.6k -44% Cost $0.0249 $0.0194 -22% ══════════════════════════════════════════════════════════════════ ``` `icm bench-agent --sessions 10 --model haiku` ### Knowledge retention Agent recalls specific facts from a dense technical document across sessions. Session 1 reads and memorizes; sessions 2+ answer 10 factual questions **without** the source text. ``` ICM Recall Benchmark (10 questions, model: haiku, 5 runs averaged) ══════════════════════════════════════════════════════════════════════ No ICM With ICM ────────────────────────────────────────────────────────────────────── Average score 5% 68% Questions passed 0/10 5/10 ══════════════════════════════════════════════════════════════════════ ``` `icm bench-recall --model haiku` ### Local LLMs (ollama) Same test with local models — pure context injection, no tool use needed. ``` Model Params No ICM With ICM Delta ───────────────────────────────────────────────────────── qwen2.5:14b 14B 4% 97% +93% mistral:7b 7B 4% 93% +89% llama3.1:8b 8B 4% 93% +89% qwen2.5:7b 7B 4% 90% +86% phi4:14b 14B 6% 79% +73% llama3.2:3b 3B 0% 76% +76% gemma2:9b 9B 4% 76% +72% qwen2.5:3b 3B 2% 58% +56% ───────────────────────────────────────────────────────── ``` `scripts/bench-ollama.sh qwen2.5:14b` ### LongMemEval (ICLR 2025) Standard academic benchmark — 500 questions across 6 memory abilities, from the [LongMemEval paper](https://arxiv.org/abs/2410.10813) (ICLR 2025). ``` LongMemEval Results — ICM (oracle variant, 500 questions) ════════════════════════════════════════════════════════════════ Category Retrieval Answer (Sonnet) ──────────────────────────────────────────────────────────────── single-session-user 100.0% 91.4% temporal-reasoning 100.0% 85.0% single-session-assistant 100.0% 83.9% multi-session 100.0% 81.2% knowledge-update 100.0% 80.8% single-session-preference 100.0% 50.0% ──────────────────────────────────────────────────────────────── OVERALL 100.0% 82.0% ════════════════════════════════════════════════════════════════ ``` - **Retrieval** = does ICM find the right information? **100% across all categories.** - **Answer** = can the LLM produce the correct answer from retrieved context? Depends on the LLM, not ICM. - The retrieval score is the ICM benchmark. The answer score reflects the downstream LLM capability. `scripts/bench-longmemeval.py --judge claude --workers 8` ### Test protocol All benchmarks use **real API calls** — no mocks, no simulated responses, no cached answers. - **Agent benchmark**: Creates a real Rust project in a tempdir. Runs N sessions with `claude -p --output-format json`. Without ICM: empty MCP config. With ICM: real MCP server + auto-extraction + context injection. - **Knowledge retention**: Uses a fictional technical document (the "Meridian Protocol"). Scores answers by keyword matching against expected facts. 120s timeout per invocation. - **Isolation**: Each run uses its own tempdir and fresh SQLite DB. No session persistence. ### Multi-agent unified memory All 17 tools share the same SQLite database. A memory stored by Claude is instantly available to Gemini, Codex, Copilot, Cursor, and every other tool. ``` ICM Multi-Agent Efficiency Benchmark (10 seeded facts, 5 CLI agents) ╔══════════════╦═══════╦══════════╦════════╦═══════════╦═══════╗ ║ Agent ║ Facts ║ Accuracy ║ Detail ║ Latency ║ Score ║ ╠══════════════╬═══════╬══════════╬════════╬═══════════╬═══════╣ ║ Claude Code ║ 10/10 ║ 100% ║ 5/5 ║ ~15s ║ 99 ║ ║ Gemini CLI ║ 10/10 ║ 100% ║ 5/5 ║ ~33s ║ 94 ║ ║ Copilot CLI ║ 10/10 ║ 100% ║ 5/5 ║ ~10s ║ 100 ║ ║ Cursor Agent ║ 10/10 ║ 100% ║ 5/5 ║ ~16s ║ 99 ║ ║ Aider ║ 10/10 ║ 100% ║ 5/5 ║ ~5s ║ 100 ║ ╠══════════════╬═══════╬══════════╬════════╬═══════════╬═══════╣ ║ AVERAGE ║ ║ ║ ║ ║ 98 ║ ╚══════════════╩═══════╩══════════╩════════╩═══════════╩═══════╝ ``` Score = 60% recall accuracy + 30% fact detail + 10% speed. **98% multi-agent efficiency.** ## Why ICM | Capability | ICM | Mem0 | Engram | AgentMemory | |-----------|:---:|:----:|:------:|:-----------:| | Tool support | **17** | SDK only | ~6-8 | ~10 | | One-command setup | `icm init` | manual SDK | manual | manual | | Hooks (auto-recall at startup) | 5 tools | none | via MCP | 1 tool | | Hybrid search (FTS5 + vector) | 30/70 weighted | vector only | FTS5 only | FTS5+vector | | Multilingual embeddings | 100+ langs (768d) | depends | none | English 384d | | Knowledge graph | Memoir system | none | none | none | | Temporal decay + consolidation | access-aware | none | basic | basic | | TUI dashboard | `icm dashboard` | none | yes | web viewer | | Auto-extraction from tool output | 3 layers, zero LLM | none | none | none | | Feedback/correction loop | `icm_feedback_*` | none | none | none | | Runtime | Rust single binary | Python | Go | Node.js | | Local-first, zero dependencies | SQLite file | cloud-first | SQLite | SQLite | | Multi-agent recall accuracy | **98%** | N/A | N/A | 95.2% | ## Documentation | Document | Description | |----------|-------------| | [Integration Guide](docs/integrations.md) | Setup for all 17 tools: Claude Code, Copilot, Cursor, Windsurf, Zed, Amp, etc. | | [Technical Architecture](docs/architecture.md) | Crate structure, search pipeline, decay model, sqlite-vec integration, testing | | [User Guide](docs/guide.md) | Installation, topic organization, consolidation, extraction, troubleshooting | | [Product Overview](docs/product.md) | Use cases, benchmarks, comparison with alternatives | ## License [Apache-2.0](LICENSE)