Stigmem v1.0: A federated knowledge fabric for AI agents ...

Cross-posted from the Stigmem blog.

Today we're releasing stigmem v1.0: A stable, open-source specification and reference implementation for a federated knowledge fabric for AI agents.

Why "stigmem"?

Stigmem = Stigmergy + Memory.

Stigmergy (Greek stigma — mark; ergon — work) is the coordination mechanism you see in ant colonies and termite mounds: agents don't communicate directly with each other. Instead, they leave traces in a shared environment: a pheromone trail, a soil deposit; and those traces guide the behavior of future agents. The colony's intelligence emerges from the environment itself, not from any central controller.

Stigmem applies the same principle to multi-agent AI systems. Agents write typed, provenance-tagged facts into a shared substrate. Other agents running later, on different platforms, inside different organizations read those facts and act on them. No central coordinator, no point-to-point protocol overhead. The knowledge environment carries the coordination signal.

The Memory half reflects persistence and decay: facts have valid_until expiries and confidence scores, so the substrate stays fresh rather than accumulating stale state just as pheromone trails fade when they're no longer reinforced.

The problem

Multi-agent systems accumulate knowledge in isolated silos. One agent knows a user prefers dark mode. Another inferred which projects are high priority. A third discovered a bug in the payment flow. None of them can see what the others know, because there's no shared place to put typed, provenance-tagged facts that travel across tool boundaries.

Stigmem gives agents that shared layer.

A fact

(entity, relation, value, source, timestamp, confidence, scope)

entity — what the fact is about (user:alice, project:payments, stigmem://org/task/42)
relation — the predicate (memory:prefers, status:blocked, infers:next_action)
value — typed payload: string, number, boolean, or JSON
source — who asserted it (an agent ID, tool name, or user session)
timestamp — Hybrid Logical Clock, monotonic across distributed nodes
confidence — 0.0–1.0, decays over time if not re-asserted
scope — access boundary (public, company, team, private)

Facts are immutable. Contradictions between nodes are surfaced as first-class conflict records, not silently overwritten.

Federation in 2 minutes

git clone https://github.com/Eidetic-Labs/stigmem && cd stigmem
docker compose up --build -d

# Register peers
docker exec stigmem-node-a-1 stigmem federation register-peer \
  --local-url http://node-a:8765 --remote-url http://node-b:8765 --scopes company,public

# Assert a fact
curl -X POST http://localhost:8765/v1/facts \
  -H 'Content-Type: application/json' \
  -d '{"entity":"user:alice","relation":"memory:prefers","value":{"type":"string","v":"dark mode"},"source":"agent:settings","confidence":1.0,"scope":"company"}'

# 30s later — fact replicated to node-b
curl 'http://localhost:8766/v1/facts?entity=user:alice&scope=company'

Two nodes start immediately. The federation handshake uses Ed25519 signatures; facts replicate under the scopes you declare. Scope enforcement is strict: private-scope facts never leave the node that created them.

MCP integration

{
  "mcpServers": {
    "stigmem": {
      "command": "npx",
      "args": ["-y", "@stigmem/mcp-server"],
      "env": { "STIGMEM_NODE_URL": "http://localhost:8765" }
    }
  }
}

Any MCP-compatible agent runtime (Claude Code, Cursor, Zed, Codex CLI) gets five tools: assert_fact, query_facts, retract_fact, synthesize_scope, and lint_scope. The synthesize_scope tool aggregates recent facts into a structured summary that slots directly into a context window so agents get fresh, scoped knowledge without managing embeddings or retrieval pipelines.

What stigmem is not

Stigmem does not replace:

Agent runtimes: it's the substrate those runtimes reason over, not a runtime itself
Orchestration platforms: the Paperclip and OpenClaw adapters emit events into stigmem; they compose, not compete
Tool protocols like MCP: MCP is the transport; the stigmem MCP adapter rides on top

It fills the gap none of them fill: typed, provenance-traceable, expiry-aware, federated shared knowledge that travels across tool and organizational boundaries.

Get involved

Stigmem is Apache 2.0 and genuinely needs contributors. A few areas where help would make the biggest difference right now:

Spec feedback: the Intent envelope (§4: goal, constraint, preference, handoff) is still in draft status and we're actively looking for real use-case feedback before stabilizing it
Adapter authors: there are stubs for additional agent runtimes and we'd love to see community-maintained adapters
Node implementations: the reference node is FastAPI + SQLite; alternative implementations in other languages are explicitly encouraged by the spec
Real-world federation topologies: if you run an interesting multi-node setup, open a discussion — we'd like to document it

The contribution process uses an RFC model: open a GitHub issue with the RFC template, discuss, then PR against the spec. New spec sections start as draft blocks and need ≥2 approvals from active contributors to merge.

If you're building something on stigmem or have questions about the federation protocol, the HLC implementation, or how to write an adapter, drop a comment or open a GitHub discussion, we read everything.

推荐订阅源

DEV Community