Claude Code 유출 이후 다중 에이전트 메모리 일관성 계층을 구축했습니다.

Engram is an MCP server that gives your agents a shared, persistent knowledge base — one that survives across sessions, syncs across engineers, and detects when two agents develop contradictory beliefs about the same codebase. Individual agent memory is solved. Engram solves what happens when multiple agents need to agree on what's true. Every agent session starts from zero. Your agent re-discovers why that architectural decision was made, re-learns which approaches already failed, re-figures out which constraints are non-negotiable. Another engineer's agent did the same thing last week. Existing memory tools fix this for a single engineer and a single agent. They don't address what happens when Agent A and Agent B — running in separate sessions, for different engineers — develop incompatible beliefs about the same system. That's a consistency problem. Engram solves it. Engram exposes four MCP tools. That's the entire surface area. | Tool | Purpose | |---|---| engram_query | Pull what your team's agents collectively know about a topic. Structured facts, ranked by relevance and recency. | engram_commit | Persist a verified discovery — a hidden side effect, a failed approach, an undocumented constraint. Append-only, timestamped, traceable. | engram_conflicts | Surface pairs of facts that semantically contradict each other. Not an error — a structured artifact. Reviewable, resolvable, auditable. | engram_resolve | Settle a disagreement. Pick a winner, merge both sides, or dismiss a false positive. | Conflict detection runs asynchronously in the background using a tiered pipeline: deterministic entity matching, NLI cross-encoder scoring, and optional LLM escalation for ambiguous cases. Commits return instantly; detection completes within seconds. - Python 3.11+ - Any MCP-compatible client — Claude Code, Cursor, Windsurf, Kiro, VS Code pip install engram-mcp engram serve Engram runs at localhost:7474 and stores facts in ~/.engram/knowledge.db . No Docker, no database setup, no API keys. Add to your MCP client config: { "mcpServers": { "engram": { "url": "http://localhost:7474/mcp" } } } Or use stdio for local-only mode: { "mcpServers": { "engram": { "command": "uvx", "args": ["engram-mcp@latest"] } } } Engram is local-first by default. To share knowledge across your team, point everyone at the same server: engram serve --host 0.0.0.0 --port 7474 Or deploy with Docker: docker run -p 7474:7474 -v engram-data:/data engram/server Every commit is immediately available to every agent on the team. ┌──────────────────────────────────────────┐ │ I/O Layer (MCP) │ ← agents connect here │ engram_commit / engram_query / │ │ engram_conflicts / engram_resolve │ ├──────────────────────────────────────────┤ │ Detection Layer │ ← runs asynchronously │ Tier 0: hash dedup + entity match │ │ Tier 1: NLI cross-encoder (local) │ │ Tier 2: numeric / temporal rules │ │ Tier 3: LLM escalation (rare) │ ├──────────────────────────────────────────┤ │ Storage Layer (SQLite) │ ← append-only, bitemporal │ facts · conflicts · agents · scopes │ └──────────────────────────────────────────┘ Every fact carries a temporal validity window (valid_from , valid_until ). Supersession, correction, archival, and versioning are all expressed through this single primitive — no pointer chasing, no separate archive tables, no decay scores. Detection is fully decoupled from the write path. The write lock is held for ~1ms (a single INSERT ). NLI inference runs in a background worker. This keeps SQLite viable under concurrent agent load. There are 400+ MCP servers that give an individual agent persistent memory across sessions. Engram is not that. Engram is a consistency layer. Other systems store and retrieve. Engram asks: are these facts coherent with each other? It is designed to be composable with existing memory tools, not to replace them. Engram is grounded in peer-reviewed research on multi-agent memory systems: - Yu et al. (2026) — frames multi-agent memory as a computer architecture problem and names consistency as the most pressing open challenge - Xu et al. (2025) — A-Mem's Zettelkasten-inspired note structure informs fact enrichment - Rasmussen et al. (2025) — Graphiti's bitemporal modeling directly inspired the temporal validity design - Hu et al. (2026) — comprehensive survey confirming shared multi-agent memory as an open frontier Full literature review: LITERATURE.md · Implementation plan: IMPLEMENTATION.md PRs welcome. See CONTRIBUTING.md for guidelines. An engram is the physical trace a memory leaves in the brain — the actual unit of stored knowledge.