Engram

Small, private, verbatim memory for your AI — set up in a minute, readable in an afternoon.

97.4% R@5 on LongMemEval, ahead of MemPalace's raw mode (96.6%), with no AI model and no API calls. Free, local, one pip install. (results)

Engram gives Claude (and any MCP client) a long-term memory that lives on your machine. It keeps every note, file and conversation word for word, organises them into wings and rooms, and finds them again with semantic search. No cloud, no API key, no account.

Why Engram

• Small enough to read. About 6,300 lines of Python in 33 files. You can audit everything that touches your data in an afternoon.

• Two commands to set up. Install, run engram install-hooks, and Claude Code saves your work and conversations automatically.

• Verbatim, not summarised. Drawers store your original text. Compression is only used for the small context Engram loads into a model.

• Private by default. Everything stays under ~/.engram/. Nothing is sent anywhere.

• Honest numbers. Benchmarks are reproducible from this repo, and unmeasured goals are labelled as goals (details).

Related MCP server: memory-mcp

Set up in 60 seconds

pip install engram-chateau                            # Python 3.9+ (or: uv tool install engram-chateau)
engram init ~/myproject                                # create your memory château
engram install-hooks                                   # auto-save from Claude Code (prints a settings.json snippet)
claude mcp add engram -- python -m engram.mcp_server   # give Claude the 22 engram_* tools

Then ask Claude to search or remember things, or use the CLI: engram mine ~/myproject --wing myapp, engram search "auth decisions".

Engram and MemPalace

Engram's design (wings, rooms, halls, drawers, a temporal knowledge graph) is inspired by MemPalace, a much larger and more mature project. Engram aims to be the small one: the same core ideas in a codebase one person can read, with a two-line setup. Here is how the two compare on LongMemEval (500 questions, session-level R@5: is the right conversation in the top 5 results?):

Engram's numbers come from one full run each on a Mac mini with the default local embedding model (all-MiniLM-L6-v2); MemPalace's are from its README. The setups differ, so read this as "level on raw retrieval":

• Engram indexes assistant messages as well as yours; MemPalace's raw mode indexes your messages only. That is why Engram answers every "what did you tell me about…" question.

• Engram's hybrid search is a keyword boost only. MemPalace's hybrid v4 also boosts by time and by your stated preferences, and leads by one point there.

• One question is worth 0.2 points.

MemPalace also publishes LoCoMo, ConvoMem and MemBench results and offers LLM reranking, a shared memory hub across machines and Docker images; Engram has none of those yet. Every Engram number above is reproducible with the commands in benchmarks/BENCHMARKS.md.

Why local memory?

Benchmark Targets

LongMemEval and ES compression are measured; the LoCoMo runner is still a stub. On LongMemEval (500 questions, all-MiniLM-L6-v2 on ChromaDB, no LLM, no reranker, no API calls) Engram's default search puts the right session in the top 5 for 97.4% of questions and in the top 10 for 98.8% (semantic search alone: 96.8% and 98.6%). See Engram and MemPalace for the comparison, and benchmarks/BENCHMARKS.md for the per-type breakdown and how to reproduce.

Quick Start

Note: Engram is published on PyPI as engram-chateau; the command and import engram are just engram. The package named engram on PyPI is an unrelated project.

# Install
pip install engram-chateau                 # or: uv tool install engram-chateau
pip install git+https://github.com/TushaeBXN/engram   # latest unreleased code from main

# Or with optional backends
pip install "engram-chateau[faiss]"         # FAISS speed backend
pip install "engram-chateau[sqlitevec]"     # zero-dependency fallback
pip install "engram-chateau[viz]"           # graph visualisation (matplotlib + networkx)
pip install "engram-chateau[all]"           # everything

# Initialise your memory château
engram init ~/myproject

# Mine a project directory
engram mine ~/myproject --wing myapp

# Mine a conversation export
engram mine ~/Downloads/claude-export --mode convos --wing myapp

# Search
engram search "auth migration decisions" --wing myapp

# Answer a question grounded in memory (uses Ollama if available)
engram answer "What did we decide about the auth strategy?" --wing myapp

# Store a typed, validated memory
engram remember-type "Always prefer float32 on CPU because bfloat16 is unsupported" \
  instruction --wing myapp --confidence 0.95

# Load cold-start context (~170 tokens)
engram wake-up

Memory Château Architecture

Wing  (person or project)
  └── Room  (named topic: "auth-migration", "ci-pipeline")
        ├── Hall  (memory type: facts | events | discoveries | preferences | advice)
        │     ├── Closet  (ES-compressed summary — fast AI read)
        │     └── Drawer  (verbatim original — never summarised)
        └── Tunnel  (cross-wing link when same room spans multiple wings)

Mined notes and conversation messages are sorted into halls by their content (e.g. "I prefer…" → preferences, "turns out the root cause…" → discoveries) using a local, rule-based classifier (engram/classifier.py). Code and config files keep their extension-based hall; when nothing matches clearly, the old extension/role default is used.

Memory Layers

Total cold-start context: ~170 tokens (L0 + L1 only).

Using with Claude via MCP

Claude Code:

claude mcp add engram -- python -m engram.mcp_server

Claude Desktop: add this to claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\):

{
  "mcpServers": {
    "engram": {
      "command": "python",
      "args": ["-m", "engram.mcp_server"]
    }
  }
}

Use the full path to the Python that has Engram installed if python on your PATH is a different one.

Claude will then have access to all 22 engram_* tools including engram_wake_up, engram_search, engram_add_memory, engram_kg_add, engram_replay, and more.

See examples/mcp_setup.md for the full tool list.

Using with Local Models

Engram works entirely offline. The vector backends use local embeddings:

from engram.backends import get_backend
from engram.palace import Palace
from engram.searcher import Searcher

palace = Palace()
backend = get_backend("chromadb")   # or "faiss" / "sqlitevec"
searcher = Searcher(backend, palace)

results = searcher.search("auth migration")
for r in results:
    print(r["text"], r["final_score"])

No model API required. ChromaDB embeds locally using its bundled models.

Full CLI Reference

Setup

engram init <dir>                              # guided onboarding + ES bootstrap

Mining

engram mine <dir>                              # mine project files
engram mine <dir> --mode convos                # mine conversation exports
engram mine <dir> --mode convos --wing myapp   # tag with a wing
engram mine <dir> --since 2026-01-01           # skip files older than date
engram mine <dir> --plugin obsidian            # use Obsidian vault plugin
engram mine <dir> --plugin notion              # Notion export
engram mine <dir> --plugin linear              # Linear issues export

Watch Mode

engram watch <dir>                             # auto-mine on file changes
engram watch <dir> --wing myapp --mode convos  # tag + conversation mode

Search

engram search "query"
engram search "query" --wing myapp
engram search "query" --room auth-migration
engram search "query" --no-decay               # disable recency weighting
engram search "query" --no-hybrid              # semantic only (keyword boost is on by default)
engram search "query" --results 20             # max results

Memory Stack

engram wake-up                                 # L0 + L1 context dump (~170 tokens)
engram wake-up --wing myapp                    # wing-scoped L1
engram wake-up --rebuild                       # rebuild L1 from drawers

Compression

engram compress                                # ES compress all closets
engram compress --wing myapp                   # wing-scoped
engram compress --wing myapp --room auth       # room-scoped

Knowledge Graph

engram kg query "Kai"
engram kg query "Kai" --all                    # include expired triples
engram kg add "Kai" works_on "Orion" --from 2025-06-01
engram kg invalidate "Kai" works_on "Orion" --ended 2026-03-01
engram kg timeline "auth-migration"

Answer (LLM-grounded QA)

engram answer "What auth approach did we choose?"
engram answer "What did we decide about the DB?" --wing myapp
engram answer "Why did we drop Redis?" --model llama3.2 --chunks 8

Typed Memory

# Store a categorised, validated memory
engram remember-type "Prefer SQLite for local dev" preference --wing myapp
engram remember-type "Use float32 on CPU because bfloat16 unsupported" \
  instruction --wing anthos --confidence 0.95

# List all 13 valid types
engram remember-type --list

Provenance

engram provenance <drawer-id>                  # provenance record for one memory
engram provenance <drawer-id> --history        # full version chain
engram provenance --summary                    # source counts across château

Graph Visualisation

pip install "engram-chateau[viz]"
engram graph                                   # all wings → memory_graph.png
engram graph --wing myapp --output myapp.png   # scoped to one wing

Maintenance

engram conflicts                               # interactive TUI conflict resolver
engram audit                                   # health check
engram audit --fix                             # auto-resolve safe issues
engram replay --room auth-migration            # chronological room story
engram replay --room auth-migration --wing myapp
engram replay --room auth-migration --type decision          # filter by type
engram replay --room auth-migration --since 2026-06-01       # filter by date
engram replay --room auth-migration --min-confidence 0.8     # filter by confidence
engram status                                  # château overview
engram split <dir>                             # split concatenated transcripts
engram split <dir> --dry-run

Configuration

~/.engram/config.json

{
  "chateau_path": "~/.engram/chateau",
  "vector_backend": "chromadb",
  "decay_factor": 0.005,
  "decay_max_days": 90,
  "hybrid_weight": 0.3,
  "collection_name": "engram_drawers",
  "people_map": {}
}

ENGRAM_VECTOR_BACKEND overrides vector_backend for a single run without editing the file, e.g. ENGRAM_VECTOR_BACKEND=sqlitevec engram search "auth".

Troubleshooting: ChromaDB fails to load

ChromaDB 1.x ships compiled Rust bindings (chromadb_rust_bindings). If they don't match your Python or platform, Engram prints a ChromaDB failed to load message instead of starting. Fix it by reinstalling (pip install --force-reinstall --no-cache-dir chromadb) or by switching to the sqlitevec backend, which needs no compiled dependencies. Each backend keeps its own index, so re-run engram mine after switching.

~/.engram/identity.txt — plain text, becomes your L0 context.

~/.engram/wing_config.json — generated by engram init.

Module Reference

Engram Shorthand (ES)

ES is a compact shorthand that any LLM can read without a decoder. It is used for the context Engram loads into a model (L1 facts, closets); drawers always keep your original text. ES is lossy: filler words such as "the" and "that" are dropped, and decompress() only expands unambiguous symbols such as ∴ and →. Expect about 1.6× on prose and 1.1× on code.

from engram.shorthand import compress, decompress

text = (
    "The authentication module is a critical component that has a dependency "
    "on the database and is responsible for verifying user credentials."
)

es = compress(text, confidence=4)
# → "auth module=★★ component +dep on db&owns: verifying user credentials. [★★★★]"

decompress(es)
# → expands unambiguous symbols (∴, →, ¬ …) back to words

# Code-aware compression
code = "def authenticate(user: str, token: str) -> bool:"
compress(code, is_code=True)
# → "fn:authenticate(user: str,token: str)->bool"

# Diff notation
compress("+add_middleware()\n-manual_verify()", is_diff=True, diff_filename="auth.py")
# → "CHANGE:auth.py add:add_middleware() rm:manual_verify()"

Recency Weighting

final_score = semantic_score × (1 + recency_boost)
recency_boost = decay_factor × max(0, decay_max_days − age_days)

Pinned drawers (engram_pin) bypass decay entirely.

Claude Code Hooks

Two hooks keep Engram up to date while you work in Claude Code:

• Save hook (PostToolUse): every file Claude writes or edits is mined into the château. Re-saving a file replaces its drawer instead of adding a duplicate.

• Pre-compact hook (PreCompact): before Claude Code compacts a conversation, the session transcript is mined so nothing is lost. Only user and assistant text is stored, not tool calls, tool output or thinking, and running it again on the same session adds no duplicates.

engram install-hooks   # copies the scripts to ~/.engram/hooks/ and prints the settings snippet

Merge the printed snippet into ~/.claude/settings.json. It looks like:

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Write|Edit|MultiEdit",
      "hooks": [{"type": "command", "command": "~/.engram/hooks/engram_save_hook.sh"}]
    }],
    "PreCompact": [{
      "hooks": [{"type": "command", "command": "~/.engram/hooks/engram_precompact_hook.sh"}]
    }]
  }
}

Optional environment variables: ENGRAM_WING (default default), ENGRAM_ROOM for transcripts (default conversation), ENGRAM_AGENT for diary entries (default claude-code).

Web UI (notes)

engram ui          # opens http://127.0.0.1:7741 in your browser

A local notes app with a force-directed graph of your notes and their [[wikilinks]]: wing tabs, #tag filters, pinned notes, a daily journal, a Markdown editor, and JSON export/import.

Note: the UI keeps its notes in your browser's local storage. It does not yet show the memories Engram stores in ~/.engram/ (mined files, conversations, MCP memories). Use engram search, engram replay or the MCP tools for those.

Requirements

• Python 3.9+

• chromadb ≥ 1.0 on Python 3.9–3.14, installed automatically. On Python 3.15+ (no onnxruntime build yet) Engram installs without it and uses the built-in sqlitevec backend; use Python 3.11–3.14 for semantic search.

• typer ≥ 0.9.0

• rich ≥ 13.0.0

• watchdog ≥ 3.0.0

• questionary ≥ 2.0.0

• pyyaml ≥ 6.0

Optional:

• faiss-cpu — FAISS backend

• sqlite-vec — sqlite-vec backend

• matplotlib + networkx — graph visualisation (pip install "engram-chateau[viz]")

• requests — Ollama API calls for engram answer

No API key. No internet after install.

Contributing

1. Fork the repo

2. pip install -e ".[dev]"

3. pre-commit install

4. Run tests: pytest tests/ -v

5. Submit a PR

Releasing

Engram is published to PyPI as engram-chateau (the command and import engram are unchanged). Publishing runs from .github/workflows/publish.yml using PyPI trusted publishing, so no API token is stored anywhere.

The trusted publisher is configured on pypi.org for project engram-chateau: owner TushaeBXN, repository engram, workflow publish.yml, environment pypi.

To release: bump version in pyproject.toml and __version__ in engram/__init__.py, merge, then publish a GitHub release tagged v<version> (for example v1.1.0). The workflow checks the tag matches both, runs the tests, builds, and uploads.

License

MIT © 2026 Tushae Thomas