Zengram

<strong>Shared memory for multi-agent AI systems.</strong>






<a href="#quick-start">Quick Start</a> &bull;
<a href="#how-it-works">How It Works</a> &bull;
<a href="#benchmarks">Benchmarks</a> &bull;
<a href="#adapters">Adapters</a> &bull;
<a href="https://github.com/ZenSystemAI/Zengram/raw/v4.5.0/docs/api-reference.md">API Docs</a> &bull;
<a href="https://github.com/ZenSystemAI/Zengram/raw/v4.5.0/docs/configuration.md">Config</a>






<a href="https://github.com/ZenSystemAI/Zengram/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/ZenSystemAI/Zengram/actions/workflows/ci.yml/badge.svg" /></a>
<a href="https://www.npmjs.com/package/@zensystemai/zengram-mcp"><img alt="npm" src="https://img.shields.io/npm/v/@zensystemai/zengram-mcp.svg" /></a>
<img alt="License: MIT" src="https://img.shields.io/badge/license-MIT-blue.svg" />
<img alt="Node 20+" src="https://img.shields.io/badge/node-20%2B-green.svg" />
<img alt="Docker" src="https://img.shields.io/badge/docker-ready-blue.svg" />
<img alt="MCP" src="https://img.shields.io/badge/MCP-compatible-purple.svg" />
<a href="https://github.com/ZenSystemAI/Zengram/stargazers"><img alt="GitHub stars" src="https://img.shields.io/github/stars/ZenSystemAI/Zengram?style=social" /></a>

Store a fact from Claude Code on your laptop, recall it from an autonomous agent on your server, get a briefing from another session — all through the same memory system. Born from a production setup where nothing existed that let multiple AI agents share memory across separate machines.

---

The Problem

You run multiple AI agents — Claude Code for development, autonomous agents for tasks. They each maintain their own context and forget everything between sessions. When one agent discovers something important, the others never learn about it.

How It Works

Typed Memory

Events are immutable history. Facts upsert by key — new facts supersede old ones. Statuses track current state. Decisions record choices and reasoning. Each type has its own lifecycle, decay rules, and mutation semantics.

Unified Storage

Every memory lives in a single Postgres database: pgvector (HNSW) for semantic vector search and structured tables for queries, entities, and full-text search. Get both "find memories similar to X" and "give me all facts with key Y" from the same system — one container, no separate vector service.

Multi-Path Search

Search runs up to three retrieval paths in parallel, fused with weighted Reciprocal Rank Fusion:

1. Vector search — Cosine similarity via pgvector (HNSW, with iterative index scans on pgvector 0.8+ so tenant-scoped queries don't lose recall)
2. Full-text search — Postgres tsvector (websearch_to_tsquery + ts_rank_cd), with an opt-in accent-folding config for mixed-language corpora (BM25_TSCONFIG=zengram_multi)
3. Entity graph (opt-in) — entities in the query expand one hop over the co-occurrence graph the write path already maintains, pulling in associatively-related memories (GRAPH_RETRIEVAL_ENABLED)

An optional cross-encoder reranker (RERANK_ENABLED) then re-scores the fused pool by reading query and document together — the highest-leverage precision upgrade in the stack (on our private bilingual production corpus it roughly doubled MRR; measure yours with the eval harness). Works with any TEI / Infinity / vLLM / Jina / Cohere-style /rerank endpoint and degrades gracefully to fused order on outage.

Final ranking blends the fused RRF signal with vector similarity (or the rerank score when enabled), then weights confidence decay, access frequency (capped), temporal proximity, and importance — items found by multiple paths genuinely rank higher. 98.4% retrieval accuracy on LongMemEval.

Built for Multi-Agent

• Cross-agent briefings — "What happened since I was last here?" returns updates from all other agents
• Cross-agent corroboration — when a second agent independently stores the same fact, it's recorded as an observation (observed_by), not dropped as a duplicate
• Credential scrubbing — API keys, JWTs, passwords, cloud/service tokens automatically redacted before storage
• Entity extraction — Regex + alias cache at write time
• LLM consolidation — Periodic background process merges duplicates, resolves contradictions, discovers connections

Benchmarks

Evaluated on LongMemEval, the academic benchmark for long-term conversational memory:

	Score
Retrieval accuracy (finding the right memories)	98.4%
QA accuracy (GPT-4o answering from retrieved context)	76.0%
Full-context GPT-4o (entire history in prompt, no retrieval)	72.4%

The benchmark uses cosine similarity only — none of the API's multi-path features (keyword full-text, temporal boost) were used. Full methodology and per-category breakdown.

LongMemEval tests single-agent chat recall. Zengram is built for multi-agent coordination — features like cross-agent briefings, typed memory, entity extraction, and credential scrubbing aren't measured by this benchmark but are core to production use.

How It Compares

Feature	Zengram	Mem0	Letta	Zep	Hindsight
Cross-machine by design	Yes	Cloud only	No	Cloud only	No
Typed memory (event/fact/status/decision)	Yes	No	No	No	No
Multi-path search (vector+full-text)	Yes	Vector only	Vector only	Hybrid	Yes
Session briefings	Yes	No	No	No	No
Credential scrubbing	Yes	No	No	No	No
Entity extraction + linking	Yes	Graph (Pro)	No	Yes	No
LLM consolidation	Yes	Inline	Self-managed	No	Reflect
Temporal validity	Yes	No	No	Yes	No
MCP server included	Yes	Community	No	No	Yes
Self-hostable (fully open)	Yes	Community ed.	Yes	Graphiti only	Yes

Quick Start

git clone https://github.com/ZenSystemAI/Zengram.git
cd Zengram

cp .env.example .env
# Edit .env — set BRAIN_API_KEY and your embedding provider key

docker compose up -d

# Verify
curl http://localhost:8084/health

# Store your first memory
curl -X POST http://localhost:8084/memory \
  -H "Content-Type: application/json" \
  -H "X-Api-Key: YOUR_KEY" \
  -d '{
    "type": "fact",
    "content": "The API uses port 8084 by default",
    "source_agent": "my-agent",
    "key": "api-default-port"
  }'

Adapters

MCP Server (Claude Code, Cursor, Windsurf)

13 tools: brain_store, brain_search, brain_briefing, brain_query, brain_stats, brain_consolidate, brain_entities, brain_delete, brain_update, brain_export, brain_import, brain_reflect, brain_research.

{
  "mcpServers": {
    "zengram": {
      "command": "node",
      "args": ["/path/to/zengram/mcp-server/src/index.js"],
      "env": {
        "BRAIN_API_URL": "http://localhost:8084",
        "BRAIN_API_KEY": "your-key"
      }
    }
  }
}

Or install via npm: npm install -g @zensystemai/zengram-mcp

Claude Code Skills

Copy adapters/claude-code/sessionend/ to your project's .claude/skills/ to get the /sessionend ritual — structured session reflections stored directly to Zengram. Full guide.

Bash CLI

• Bash: ./adapters/bash/brain.sh store --type fact --content "Server migrated"
• Any HTTP client: Plain REST — full reference

Documentation

Doc	Description
API Reference	Every endpoint with request/response examples
Architecture	System design, data flows, component inventory
Configuration	All environment variables
Data Model	Memory types, decay, dedup, supersedes logic
MCP Tools	The 13 MCP tools agents use
Operations	Deployment, monitoring, failure modes
Benchmarks	Full LongMemEval methodology and results
Examples	curl demo, Python client

Roadmap

Recently shipped: cross-encoder reranking stage, entity-graph retrieval path, weighted RRF, self-hosted encoder support (local endpoints + instruction prefixes + in-place re-embed), agentic iterate-until-sufficient retrieval (brain_research) with grounded [mem:<id>] citations, pgvector migration (single-Postgres storage), multi-collection support, on-demand LLM reflection, temporal validity — full changelog

Coming next: Automatic memory capture, hosted docs, LangChain/LlamaIndex integration

Contributing

Contributions welcome! See CONTRIBUTING.md.

See Also

• OpenClaw Memory Toolkit — Long-term memory for OpenClaw agents.

License

MIT — see LICENSE.

---

Built by ZenSystem AI