The Problem

Every time you start a new AI session, your assistant has zero memory of what you built yesterday. The architecture decisions, the bugs you fixed, the preferences you stated — all gone. You spend the first 10 minutes re-explaining context.

The Fix

context-mem runs in the background, captures everything automatically, and retrieves exactly the right context when you need it:

• Longer sessions without losing context (99% token savings)
• Instant continuity — new sessions pick up where you left off
• Automatic — no manual saving, no commands to remember
• Fully local — your code never leaves your machine
• Free — no API keys, no subscription, no cloud

npm i context-mem && npx context-mem init

One command. Works with Claude Code, Cursor, Windsurf, VS Code, Cline, and Roo Code.

Retrieval Benchmarks

Tested on 4 academic benchmarks. All scores are session-level retrieval recall (did the correct session appear in top-k?), not end-to-end QA accuracy.

Pure Local (zero API calls, fully free)

With Optional LLM Reranking (Haiku, ~$1 per 500 queries)

vs MemPalace (same methodology — session-level retrieval recall)

Both systems achieve 100% on LME with optional Haiku reranking. MemPalace comparison uses identical methodology (session-level, same datasets).

How It Works

Every tool output flows through the pipeline: privacy screening (9 secret detectors) → parallel extraction (entities, importance, topics) → 14 content summarizers → triple storage (verbatim archive, SQLite summaries, knowledge graph) → adaptive compression over time.

Full coding session (50 tool outputs): 365 KB → 3.2 KB (99% savings).

What it is (and isn't)

context-mem is:

• A retrieval-first memory system (not a chatbot wrapper)
• A context compression engine (14 content-aware summarizers)
• Infrastructure for AI agents (44 MCP tools)

context-mem is not:

• Chat history storage (it extracts meaning, not raw logs)
• An LLM wrapper (works without any API keys)
• A cloud service (fully local SQLite)

Quick Start

npm i context-mem && npx context-mem init

init auto-detects your editor:

Real-World Examples

You: "Why did we choose Postgres?"
  → recall returns the exact verbatim quote from March 15, importance 0.95,
    with the full evidence chain: error → file_read → search → decision

You: "What did Sarah work on last sprint?"
  → browse by person shows 14 observations mentioning Sarah,
    grouped by topic (auth, database, deployment)

You: "Generate a PR description"
  → context-mem story --format pr assembles changes, decisions, resolved
    issues, and test plan from the current session

You: "What are we about to forget?"
  → predict_loss shows 8 entries at risk: low importance, 45+ days old,
    never accessed. Pin the critical ones before they decay.

Search Architecture

BM25 (8 strategies + synonym expansion) and vector search run independently in parallel, then fuse via intent-adaptive weights with IDF-weighted content reranking. Optional LLM judge reranker pushes accuracy to 100%. Fully local by default.

Core Features

Intelligence Dashboard

Real-time web UI with 6 pages — context-mem dashboard to launch:

More dashboard pages

Knowledge Graph — force-directed entity visualization with type filtering and depth control:

Topics — topic cloud with observation counts and cross-project tunnels:

Timeline — chronological observations with importance badges, flags, and verbatim mode:

How It Compares

Performance

All operations are sub-millisecond, zero LLM dependency:

MCP Tools (44)

CLI Commands

context-mem init                    # Initialize in current project
context-mem serve                   # Start MCP server (stdio)
context-mem status                  # Show database stats
context-mem doctor                  # Run health checks
context-mem dashboard               # Open web dashboard (6 pages)
context-mem why <query>             # Decision trail — why was X decided?
context-mem story --format pr       # Generate narrative (pr/standup/adr/onboarding)
context-mem import-convos <path>    # Import conversations (auto-detect format)
context-mem export                  # Export as JSON
context-mem import                  # Import from JSON
context-mem plugin add|remove|list  # Manage summarizer plugins

Configuration

{
  "storage": "auto",
  "plugins": {
    "summarizers": ["shell", "json", "error", "log", "code"],
    "search": ["bm25", "trigram", "vector"],
    "runtimes": ["javascript", "python"]
  },
  "search_weights": { "bm25": 0.45, "trigram": 0.15, "levenshtein": 0.05, "vector": 0.35 },
  "privacy": { "strip_tags": true, "redact_patterns": [] },
  "lifecycle": { "ttl_days": 30, "max_db_size_mb": 500, "max_observations": 50000 },
  "ai_curation": { "enabled": false, "provider": "auto" }
}

Platform Support

Documentation

License

MIT — Juba Kitiashvili

npm i context-mem && npx context-mem init