icm

English | Français | Español | Deutsch | Italiano | Português | Nederlands | Polski | Русский | 日本語 | 中文 | العربية | 한국어

Tell Claude Code how your project handles auth on Monday, and Tuesday's Gemini CLI session already knows it. ICM keeps what your coding agents learn (decisions, fixes, conventions, preferences) in one SQLite file on your machine, and gives the relevant part back at the start of each session and with each prompt you send. Up to 18 agents and editors share that memory, so you stop re-explaining your project every time you open a session or switch tools.

• 92.9% on LoCoMo (1,540 questions, mean of three runs), level with Hindsight (92.0%), with a third less context per question. Details and caveats below.
• No LLM call to store or recall. Hindsight, Mem0, Graphiti (Zep) and claude-mem call an LLM for every memory they store by default. ICM does not. Only its automatic extraction of facts from tool output goes through the LLM command-line tool you already use, when one is installed; provider = "none" keeps that local too (see Quickstart).
• Useful without an embedding model. Keyword recall alone puts at least one of the right sessions in the top 5 for 88.6% of LoCoMo questions, with a median of 6.8 ms per recall on Linux x86-64.
• 97.4% on LongMemEval-S, retrieval with no LLM: one of the right sessions in the top 5 for 487 of 500 questions, against 96.6% for MemPalace and 95.2% for agentmemory on the same measure.
• Not ahead everywhere. On PersonaMem (a user's evolving preferences), Hindsight leads: 86.6% against ICM's 81.7%.

Quickstart

brew tap rtk-ai/tap && brew install icm    # or: curl -fsSL https://raw.githubusercontent.com/rtk-ai/icm/main/install.sh | sh
icm init                                   # instructions, skills and hooks for every agent it detects

That is the whole setup. Open a new session in Claude Code, Codex, Gemini CLI or Copilot CLI: your agent now starts with a short pack of the most important memories of the project it is in (those whose topic carries the repository's name, such as decisions-myapp in a repository called myapp, plus your preferences) and receives the memories relevant to each prompt you send. What it learns from its tool output is queued, and the queue is turned into memories at the end of each Claude Code session; with the other tools, run icm extract-pending (from a cron job, for instance).

Storing and recalling stay on your machine. The automatic extraction hands text to the LLM command-line tool you already use (Claude Code, Codex or Gemini CLI) when one is installed; set provider = "none" under [extraction.summarizer] to keep it entirely local.

To see it work right away, store and recall by hand:

$ icm store -t decisions-myapp -c "Auth uses short-lived JWTs, refreshed through /auth/refresh" -i high
Stored: 01M47YY6CHVZ48BYKQRCRNZTCF

$ icm recall "how does auth work"
memories[1]{id,topic,importance,weight,summary}:
  01M47YY6CHVZ48BYKQRCRNZTCF,decisions-myapp,high,0.975,"Auth uses short-lived JWTs, refreshed through /auth/refresh"

The first store or recall with semantic search downloads the multilingual embedding model once (Qdrant/multilingual-e5-large-onnx, about 2 GB). To try ICM without it, add --no-embeddings (keyword recall, as in the output above) or pick a lighter model in the config. icm init writes hooks and instructions into each detected agent's configuration; icm uninstall --dry-run shows how to remove them. Windows, Linux, Nix and building from source: Install.

Benchmark comparison

Answer accuracy on LoCoMo (10 long conversations, 1,540 questions), measured with the public Agent Memory Benchmark harness: the memory system retrieves context, gemini-3.1-pro-preview answers from it, gemini-2.5-flash-lite judges the answer.

On PersonaMem 32k (589 multiple-choice questions about a user's evolving preferences, same harness and answering model, scored by letter match):

Retrieval alone, on LoCoMo, with no answering model: the share of questions for which at least one of the right sessions is in the top results. This is the evidence for the recall engine itself.

Both engines were measured with the same 0.11 build. The 0.11 runs received each session's date and the question's date; the 0.10 engine has no date input, so its runs received none.

LongMemEval-S, retrieval only, with no LLM (ICM with its default embedding model; 500 questions; each question comes with about 48 past sessions to search; one memory per session, user turns only, as MemPalace indexes them; no date given to ICM):

MemPalace and agentmemory figures are recomputed with our scorer from the result files each project publishes; they match their published numbers. agentmemory indexes all turns of a session; on that unit BM25 alone reaches 96.2% and 83.0%. A plain BM25 already scores in the mid-nineties on the first measure, which is why the second one, all the right sessions, separates the systems better.

What these numbers do and do not show:

• ICM and Hindsight are tied on LoCoMo. ICM's three runs (92.9%, 93.1%, 92.9%) are each about 1 point above Hindsight's published 92.0%, a gap of about 14 questions, inside the sampling error (95% interval for ICM: 91.6 to 94.2). ICM gets there with a third less context and without calling an LLM when a memory is stored.
• On PersonaMem, Hindsight is ahead by 4.9 points, outside the sampling error (95% interval for ICM: 78.6 to 84.8). ICM is also 2.7 points below the hybrid search baseline, inside that interval, while reading a third less context than it. The three runs spread from 80.1% to 82.5%.
• Not identical conditions. The harness is maintained by Vectorize, the vendor of Hindsight. The published LoCoMo results predate a change that set the answer and judge temperature to 0; our run uses the current harness (commit f618ed7) and Vertex AI.
• At 50 chunks, much of each conversation is returned, so this accuracy also measures the answering model. The retrieval table is the evidence for the recall engine itself.
• Three runs per dataset. The answering model varies from run to run: 0.2 points on LoCoMo, 2.4 points on PersonaMem. The intervals above cover the sampling of questions.

• By question type (LoCoMo, harness labels, three runs): open-domain 96.7% (841 questions), temporal 90.9% (321), single-hop 88.8% (282), multi-hop 78.8% (96).
• Latency. Median recall latency 137 to 149 ms on the cluster's 4-vCPU nodes during the LoCoMo runs; 272 sessions ingested with no LLM call. In the retrieval-only runs on Linux x86-64: median 136 ms with embeddings, 6.8 ms keyword-only.
• Session dates. The benchmark adapter writes each session's date into ICM's memory text; Hindsight receives the same dates as metadata, and every system gets the date of the question.
• **The weakest category is the 96 questions the harn