Zengram

What Zengram does

Zengram stores everything in one Postgres database, using pgvector with HNSW for semantic search alongside structured tables for entity and full-text queries. Memories are typed: events are immutable history, facts upsert by key so newer facts supersede older ones, statuses track current state, and decisions record choices and reasoning. Search runs up to three paths in parallel, vector, Postgres full-text and an opt-in entity graph, fused with weighted Reciprocal Rank Fusion, with an optional cross-encoder reranker.

For multi-agent use, it offers cross-agent briefings that answer what happened since an agent was last active, records a second agent's matching fact as an observation instead of dropping it as a duplicate, redacts API keys and tokens before storage, and runs periodic LLM consolidation. The README reports 98.4% retrieval accuracy and 76.0% QA accuracy on LongMemEval using cosine similarity only, and notes that the benchmark tests single-agent chat recall. GitHub topics tag the repository with hermes-agent, openclaw, claude-code, codex, cursor and gemini.

Key features

• Typed memory: events, upserted facts, statuses and decisions
• Single Postgres database with pgvector HNSW and structured tables
• Up to three parallel search paths fused with weighted Reciprocal Rank Fusion
• Cross-agent briefings and corroboration recorded as observed_by
• Automatic redaction of API keys, JWTs and passwords before storage
• Periodic LLM consolidation that merges duplicates and resolves contradictions

When to use it

• Sharing facts between Claude Code on a laptop and an autonomous agent on a server
• Getting a briefing of what other agents did since your last session
• Keeping one source of truth where newer facts replace older ones

Who it is for: People running several AI agents on separate machines who want them to share one memory.

How it fits with Hermes Agent

Supports Hermes Agent among other agents: the repository is tagged hermes-agent alongside openclaw, claude-code, codex, cursor and gemini, and exposes its memory through MCP.

FAQ