remem-mcp

Your coding agent stops repeating the same mistakes.

Local memory that survives context compaction — learns from every error, injects fixes before the next attempt, and syncs to your git repo so your whole team shares it.

Install

npx remem-mcp setup

That's it. Auto-detects Claude Code, Cursor, Devin, Codex. Registers MCP server + hooks. Restart your agent.

npx remem-mcp demo     # Live demo: real build, real errors, real hooks
npx remem-mcp demo-codegraph  # Live CodeGraph demo on facebook/react
npx remem-mcp status   # One dashboard: everything at a glance

The demo creates a real TypeScript project, runs real npm run build, captures real TS2307 errors, and shows the full learning loop — capture → inject → fix → upvote → cross-project inheritance. No hardcoded strings.

Related MCP server: knitbrain

Why it's different

Per-agent install

claude mcp add remem-mcp --scope user -- npx -y remem-mcp
npx remem-mcp install-hooks

Or add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "remem-mcp": { "command": "npx", "args": ["-y", "remem-mcp"] }
  }
}

devin mcp add remem-mcp --scope user -- npx -y remem-mcp
npx remem-mcp install-hooks

Add to ~/.codex/config.toml:

[mcp_servers.remem-mcp]
command = "npx"
args = ["-y", "remem-mcp"]

[mcp_servers.remem-mcp.env]
TDAI_GLOBAL_SESSION_KEY = "global"

Then run npx remem-mcp install-hooks.

MCP tools require sandbox_mode = "danger-full-access".

How it works

Memory lives in a local SQLite database — outside the agent's context window. When the agent compacts or starts a new session, memory is re-injected automatically. No more re-explaining what you already told it yesterday.

PreCompact hook: when the agent is about to compact context, remem-mcp saves a checkpoint (decisions made, approaches tried, what's verified working) to the DB. After compaction, the agent recalls it — so the compact doesn't destroy your session's learnings.

Two layers: automatic (runs via hooks, zero tool calls) and on-demand (you call when you need deeper context).

Automatic — three learning loops + compaction survival

All run via lifecycle hooks. The agent doesn't need to call any tool.

1. Error learning — command fails → capture → inject fix before next attempt → succeed → upvote.

2. Decision learning — npm install, git commit, config → auto-capture → inject past decisions before similar commands.

3. Pattern learning — Write/Edit → auto-capture code patterns → inject same-language patterns before editing.

4. Compaction survival — PreCompact hook fires before context compaction → saves checkpoint → agent recalls after compact. Memory survives.

On-demand — CodeGraph, Wiki, Search

When the automatic loops aren't enough, use these for deeper code navigation.

# 1. Index your codebase (one-time, rerun after major changes)
npx remem-mcp index --path src --repo .

# 2. Search symbols (auto-scoped to current directory)
npx remem-mcp search-code --query "parseTar"
# → parseTar  at  src/parse.ts:22

# 3. List symbols in a file
npx remem-mcp list-code src/reporters/fancy.ts
# → Class    L49-135  FancyReporter
# → Method   L86-134  formatLogObj

# 4. Trace callers / callees / impact (use symbol ID from step 2)
npx remem-mcp callers 01KZXPPHF93TS4HV8FWCSSK36A
npx remem-mcp impact  01KZXPPHF93TS4HV8FWCSSK36A

# Wiki + viewer
npx remem-mcp wiki ingest --path docs      # Index markdown docs + ADRs
npx remem-mcp wiki outdated                 # Find outdated wiki pages
npx remem-mcp viewer                        # Web UI at localhost:7331

• CodeGraph — symbol search, callers/callees, impact analysis. Auto-scoped to your project — no cross-project contamination.

• Wiki — markdown docs, ADRs, outdated detection.

• Search — hybrid BM25 + sqlite-vec vector search with RRF fusion. explain_recall shows scores.

Daily commands

npx remem-mcp status           # Everything at a glance
npx remem-mcp viewer           # Web UI at localhost:7331
npx remem-mcp errors           # Error dashboard
npx remem-mcp decisions        # Decision dashboard
npx remem-mcp patterns         # Pattern dashboard
npx remem-mcp recent [N]       # Recent captures
npx remem-mcp help all         # Full list of 40+ subcommands

Configuration

All settings have defaults. Config file is optional: ~/.config/remem-mcp/config.json.

Team sharing — npx remem-mcp sync-export writes .remem-mcp/memory-export.jsonl. Commit it to git. Team members get the same memory on git pull (auto-imports on startup).

TypeScript SDK

import { Memory } from "remem-mcp";

const memory = new Memory();
await memory.capture("We chose SQLite for storage.", "decision", ["arch"]);
const results = await memory.recall("storage decision");

Benchmark

remem-mcp is evaluated against the same benchmarks as TencentDB Agent Memory, plus the Agent Memory Benchmark (AMB) suite.

• PersonaMem — bowen-upenn/PersonaMem (588 questions, 20 personas, multiple-choice QA). TencentDB reports 76% with memory enabled, 48% without. remem-mcp scores 80% using a search-recall proxy (no LLM API key needed).

• LoCoMo — long conversation multi-hop QA (19 sessions, 400+ turns). remem-mcp scores 85% with keyword-heuristic scoring.

• AMB — Agent Memory Benchmark (L1: 56 recall tests, L2: 5 multi-session scenarios, L3: 1K+ memories with distractors). remem-mcp scores 100/100/100.

• LongMemEval — xiaowu0162/LongMemEval (ICLR 2025, 500 questions, 5 memory abilities: temporal reasoning, multi-session, knowledge update, single-session recall, abstention). remem-mcp scores 92% on the oracle variant.

Run the benchmarks:

bash scripts/bench-all.sh           # Full: AMB + LoCoMo + PersonaMem (~5 min)
bash scripts/bench-all.sh --quick   # AMB only (~2 min)

Credits

Core based on TencentDB Agent Memory (MIT, Tencent 2026). Replaces the cloud backend with embedded SQLite + sqlite-vec + FTS5. Adds error/decision/pattern learning loops and lifecycle hooks.

License

MIT. See LICENSE.