Engram

Every assistant starts from zero. Tell Claude you're job hunting, and ChatGPT still doesn't know. Ask any of them "how is my job search going?" and the best a vector store can do is return the five sentences that look most similar — it can't tell you that you've applied 14 times in 16 days, that 3 became interviews, or that system design came up weak in both onsites.

Engram is a single, self-hosted memory service that every assistant shares. Agents call remember when something happens and recall(situation) before they act. Engram turns free text into typed memories in the background and answers with one ranked, token-budgeted brief:

An illustrative brief:

recall("I'm applying for jobs")

## You
- Backend/AI engineer, prefers concise answers, targets remote roles

## Current task
- deadline: CV to Acme by Friday

## Job search so far
14 events since 2026-09-14 (16d), last 2026-09-30 · applied 10 · interview 3 · rejection 1 · outcomes: rejected 1
Momentum is good; DSA rounds go well, system design is the recurring weak spot.

## Lessons
- [r:7] System design was flagged in 2/2 onsite interviews — prep it before the next one

## Relevant history
- [e:41] 2026-09-29 Acme onsite: system design round went poorly
- [e:38] 2026-09-24 Beta Labs rejection after recruiter screen

## How you do it
- [f:2c…] How to apply: 1. cv send <company> → 2. log it → 3. follow up in 7 days

How it works

Engram uses every kind of memory where it fits, and one retrieval pipeline ranks them together.

Everything is joined by entities (topic:job-search, company:acme, project:cityfix): mention one in a situation and every layer that touches it comes along.

flowchart LR
    A["Any assistant<br/>(MCP / REST / hooks)"] -- remember --> Q[(ingest queue)]
    Q --> X["extract<br/>1 LLM call → typed JSON"]
    X --> R["resolve entities<br/>split · dedupe · redact"]
    R --> E[(episodes)]
    R --> F[(mem0 facts &<br/>procedures)]
    R --> S[(session state)]
    F --> C{"reconcile<br/>duplicate / supersedes"}
    E & F --> Z["consolidate (sleep pass)<br/>digests · reflections · profile"]
    A -- "recall(situation)" --> P["plan<br/>entities · weights"]
    P --> CH["channels: profile · session · aggregates ·<br/>episodes · facts · procedures · lessons · docs"]
    CH --> K["rank: RRF × importance × recency × confidence"]
    K --> B["token-budgeted brief"] --> A

Write path. remember returns immediately; a worker extracts entities, events, facts, procedures and session notes in one LLM call, resolves dates in your timezone, redacts secrets, splits multi-company career events so counts stay right, de-duplicates events that were mentioned twice, and reconciles new facts against old ones (a changed preference supersedes the old one — hidden from recall, kept in history).

Read path. recall matches entities (aliases, embeddings, one hop of co-occurrence), runs the channels, fuses them with reciprocal-rank fusion, applies importance, recency half-lives and confidence, and packs a sectioned brief under your token budget. fast=true uses no LLM at all, and recall keeps working (profile, counts, full-text) even if the embedding model is down.

Sleep pass. Every 6 hours (and after sessions end) Engram rebuilds topic digests, derives lessons from ≥2 related events (with the events as evidence), refreshes your profile and expires old session notes. Nothing is deleted.

Full design: docs/design.md.

Related MCP server: mnemosis-mcp

Quick start

Requirements: Python 3.11+ with uv, Postgres 16 with pgvector, and Ollama for local embeddings.

git clone https://github.com/mohammed-usmani/engram && cd engram
uv sync
ollama pull nomic-embed-text            # embeddings run locally and free

cat > .env <<'ENV'
CONNECTION_STRING=postgresql+asyncpg://postgres:postgres@localhost:5432/engram
DATA_DIR=data/examples                  # text files imported once on first start (optional)
USER_NAME=Alex                          # how prompts refer to you
USER_TZ=Asia/Kolkata                    # "yesterday" means *your* yesterday
TOGETHER_API_KEY=...                    # or GEMINI_/GROQ_/MISTRAL_/OPENAI_/ANTHROPIC_API_KEY
MEMORY_LLM_CHAIN=together,ollama        # tried in order; local Ollama as the fallback
MEMORY_MODEL_TOGETHER=deepseek-ai/DeepSeek-V4-Flash-0731
ENV

uv run alembic upgrade head
uv run uvicorn main:app --host 127.0.0.1 --port 8001

On macOS, scripts/launchd/install.sh installs it as a login service (plus nightly backups). On Linux, scripts/systemd/install.sh does the same with systemd user units (sudo loginctl enable-linger $USER keeps it running while logged out).

Try it:

curl -X POST localhost:8001/api/memory/remember -H 'Content-Type: application/json' \
  -d '{"text": "Applied to Zomato and Swiggy for SDE-2 roles last Monday", "agent": "curl"}'
# a few seconds later
curl -X POST localhost:8001/api/memory/recall -H 'Content-Type: application/json' \
  -d '{"situation": "how is my job search going?", "fast": true}'

Connect your assistants

Conversations save themselves. scripts/transcript_sync.py (installed by the launchd/systemd scripts, every 10 minutes) reads Claude Code, Codex CLI and Gemini CLI transcripts and sends new user/assistant prose to Engram in chunks, each stamped with the agent and the time of its last message. Long and still-open sessions are captured in full; one offset file shared with the hook means nothing is sent twice. Tools whose transcripts aren't readable (Antigravity, web apps) save through remember.

Every write says who and when. remember requires agent (claude-code, claude-web, chatgpt, gemini, codex…) and occurred_at (when it happened, ISO 8601); teach and the document tools require agent, and documents record it as updated_by. Writes without them are rejected.

Local tools on http://localhost:8001/mcp (Claude Code, Codex, Gemini CLI, Antigravity, Cursor) need no token or sign-in; the token and OAuth only apply to requests that arrive through a tunnel.

For assistants without hooks, add one custom instruction: "Before answering anything personal or starting a task, call recall with my words. When I share a fact, preference, decision or outcome, call remember."

Tools: recall · remember · note · teach · expand · timeline · forget, and for documents list_documents · get_document · search_context · edit_document · save_document · delete_document

Your documents

Resume, projects, skills, experience, education, achievements, certifications and profiles live in the database and are edited in the browser at http://localhost:8001/admin, or by any assistant over MCP (edit_document for a small change, save_document to create or replace) — no files, no reseeding. Write plain text: Key: value lines at the top become fields, a line like Summary: starts a section; both are re-derived and the document is re-embedded on every save, so recall sees the change immediately. Documents appear in briefs under Documents (e.g. [d:resume/master_resume]).

A profile document (profile/linkedin, profile/indeed, profile/github…) holds the exact text a public profile currently shows, so an assistant asked "what does my LinkedIn say?" quotes it instead of guessing. Note anything you haven't captured yet in the document itself, so nothing gets invented.

The admin pages follow the same rule as everything else: no token from your own browser on localhost, the token for anything arriving through a tunnel.

Text files are only an import path: on an empty database Engram imports DATA_DIR (data/examples/ ships a fictional sample), and Import new files adds files that aren't in the database yet. Importing never overwrites a document you've edited.

See what it knows

Open http://localhost:8001/memory for a browser view of the whole memory:

Anything wrong? Forget removes it everywhere at once — including the profile at the top of every brief, which is built live from your current facts.

Choosing the extraction model

Every memory write costs one LLM call (plus a small one when a new fact may contradict an old one). Any OpenAI-compatible provider works: Together, Groq, Mistral, Gemini, Cerebras, OpenAI, Anthropic, or local Ollama. Pick one with the included eval — 14 labelled cases (multi-company applications, relative dates, interview + offer in one sentence, procedures, session notes, noisy coding transcripts, small talk that must be ignored), repeated to catch flakiness:

uv run python scripts/eval_extraction.py --chain together --together-model <model> --repeats 5
uv run python scripts/eval_extraction.py --chain ollama --ollama-model qwen2.5-coder:7b --concurrency 1

Results at the time of writing:

*≈ 15 coding sessions, 20 notes and 30 recalls a day. Writes are asynchronous, so latency never blocks your assistant.

Batch mode for big backlogs

When the queue piles up (an import, a long day of sessions), the Queue tab on the Memory page has a Switch queue to batch mode button. The live workers then stand aside and, about once a minute, every waiting job goes to Together's Batch API as one batch; results come back in minutes to hours and are saved exactly like live extractions. Batches already sent are still collected after you switch back. Batches cost about half and use their own rate limits, but Together only batches serverless models (DeepSeek-V4-Flash is refused), so they use MEMORY_BATCH_MODEL, default openai/gpt-oss-120b (126/126 on the eval above).

Your data stays yours

• Local-first. One process on 127.0.0.1:8001; embeddings run on your machine.

• Secrets are redacted before anything is stored — API keys, tokens, private keys, URL credentials, card numbers (Luhn-checked), "my password is …".

• Auth. With ADMIN_TOKEN set, every path except /api/health (and the OAuth endpoints below) requires the bearer token for remote or tunnelled requests; direct local use keeps working.

• OAuth for apps that can't send a header (ChatGPT, Gemini). Set PUBLIC_URL to the HTTPS address the app reaches (e.g. your Tailscale Funnel URL) alongside ADMIN_TOKEN. Engram then serves MCP OAuth discovery, client registration, /authorize, /token and /revoke. Connecting sends you to /oauth/consent, where you approve by typing ADMIN_TOKEN, so OAuth never grants more than the token does. Clients and token hashes are kept in OAUTH_STORE (default ~/.config/engram/oauth.json); access tokens last an hour and refresh silently, refresh tokens 180 days. Delete that file to sign every app out.

• Encrypted backups. scripts/backup.sh dumps Postgres + your documents + fact history into one age-encrypted archive (nightly via launchd, keeps 14). Point ENGRAM_BACKUP_DIR at a private git repo and every backup is pushed off-machine. scripts/restore.sh rebuilds everything on a new machine and refuses to overwrite a database that already holds memories unless you pass --force.

• Several devices. Run Engram on one machine and connect the others over Tailscale with ADMIN_TOKEN — one memory, no sync conflicts.

API

Engram also ships a document manager (/admin), a document store with hybrid search, and a RAG chat (/chat) over your documents.

Development

createdb engram_test   # TEST_DATABASE_URL defaults to <CONNECTION_STRING db>_test
uv run pytest

src/memory/      models · llm (provider chain) · facts (mem0) · entities · extract · ingest
                 recall · consolidate · worker · api · dashboard (/memory) · backfill
src/mcp_server.py  MCP tools (also served at /mcp by main.py)
scripts/         claude_memory_hook.py · eval_extraction.py · backup.sh · restore.sh · launchd/
data/examples/   fictional sample documents
docs/design.md   design notes

License

MIT © 2026 Mohammed Usmani