Skip to main content
Version: v0.9.0a2

Memory Garden

4 min readEvaluator · Integrator · CuratorPublic identity

Memory Garden is Stigmem's public identity: an open, federated knowledge fabric where AI agents and humans store facts that travel — across tools, runtimes, companies, and deployments — with full provenance and zero vendor lock-in.

The problem with closed memory

Every major agent memory product today — mem0, Letta, Zep — is a closed pool. Your agent's memory lives in one vendor's infrastructure, speaks that vendor's API, and stays inside that vendor's boundaries.

No shared brain across frameworks

When you add a second agent framework, you start from scratch.

Memory stays behind at boundaries

When you cross a team or company boundary, memory stays behind.

Vendor hostage

When the vendor changes their API or pricing, your memory is hostage.

No audit trail

When an agent makes a decision, there is no audit trail linking it to the facts it relied on.

These are not edge cases. They are the daily frictions of running any multi-agent system at scale.

What Memory Garden does differently

Memory Garden stores every fact as an immutable, typed record:

(entity, relation, value, source, timestamp, confidence, scope)
Field
Carries
What it does
entity
subject
What the fact is about — user:alice, company:acme, project:loom.
relation
predicate
What kind of claim this is — memory:role, roadmap:status, preference:timezone.
value
object
The typed value: string, number, boolean, datetime, or a reference to external content.
source
provenance
Who asserted it — agent:ceo, user:alice, system:intake.
timestamp
when
Immutable, set by the node.
confidence
certainty
1.0 = certain, 0.5 = uncertain, 0.0 = retracted.
scope
visibility
local, team, company, or public (federatable).

Every fact is written once and never mutated.

Updates are new facts. The latest fact for an (entity, relation, scope) triple wins — unless two sources contradict each other, in which case both facts are surfaced and the contradiction is flagged explicitly for resolution.

How it differs from the alternatives

Capability
Memory Garden
mem0 · Letta · Zep
License
Apache 2.0
Proprietary SaaS · Proprietary · Proprietary SaaS
Self-hostable
Yes — single binary
No (enterprise tier) · Partial · Yes (complex)
Federated
Yes — signed handshakes
No · No · No
Provenance on every fact
Yes
No · No · Temporal edges only
Contradictions surfaced
Yes — first-class conflict records
No · No · No
Entity-scoped facts
Yes — facts belong to entities, not agents
No · No · No
Open spec
Yes — Apache 2.0, community namespace registry
No · No · No
Platform-agnostic
Yes — MCP, OpenClaw, Paperclip adapters
Custom SDK · Letta-native · Zep-native

Entity-scoped vs. agent-scoped is the most important architectural difference.

In every existing memory system, the CEO's memory about project:loom and the CTO's memory about project:loom are separate stores with no reconciliation path. In Memory Garden, those facts live in the same entity namespace — any authorized agent queries and contributes to the same set of facts about the same entity. There is one ground truth, not N copies of it.

The Open Connector Mesh

Memory Garden ships with an open connector mesh: a set of adapters and integrations that let agents read from and write to the fabric without per-agent plumbing.

Platform adapters

Adapter
Connects to
What it does
MCP adapter
Claude Code agents
Exposes stigmem_assert and stigmem_query as MCP tools — any Claude Code agent with the server in .mcp.json gets full read/write access.
Paperclip adapter
Paperclip workflow
Wires into Paperclip's hook system to emit issue lifecycle events (checkout, status changes, blockers) as facts automatically.
OpenClaw adapter
PARA MEMORY.md
Drop-in replacement for PARA MEMORY.md reads and writes, routing to a Memory Garden node.

Business integrations

Connectors run bidirectionally. Agents don't just read from external systems; they can write decisions back as facts into the fabric, so the next agent or human that queries the same entity sees the full history.

Shopify

Product catalog, order events, customer state.

Stripe

Billing events, subscription state, payment status.

GitHub

PR status, issue events, CI results.

AWS

Infrastructure events, deployment facts.

Google Workspace

Calendar, Drive, Gmail signals.

WooCommerce

Order and inventory state.

Federated by design

Memory Garden nodes can peer with each other using a signed handshake protocol. Two nodes exchange PeerDeclaration documents — signed with Ed25519 keys — specifying exactly which scopes they share and in which direction.

Scope
Federation behavior
Notes
local · team
never leave origin node
No federation under any circumstances.
company
cross only via explicit peer declaration
The active peer declaration must include it.
public
replicates by default
Via cursor-based pull replication.

Provenance is preserved across federation.

A fact replicated from another node retains its original source, timestamp, and confidence. The relay chain is transparent. An agent on Node B can always see that a fact originated from agent:ceo on Node A — not from Node A itself.

This is the same model as email federation (SMTP), the Fediverse (ActivityPub), and personal data pods (Solid): every node is self-hosted and sovereign. There is no central registry, no central operator, no single point that can be taken down or monetized behind a paywall.

Getting started

# Run a Memory Garden node locally
pip install stigmem-node
stigmem-node

# Assert your first fact
curl -X POST http://localhost:8765/v1/facts \
-H "Content-Type: application/json" \
-d '{
"entity": "user:alice",
"relation": "memory:role",
"value": { "type": "string", "v": "Lead Engineer" },
"source": "agent:onboarding",
"confidence": 1.0,
"scope": "company"
}'

# Query it back
curl "http://localhost:8765/v1/facts?entity=user:alice&relation=memory:role"

See the Quickstart guide for a full walkthrough including the MCP adapter setup and your first federated peer.

What Memory Garden is not

Not a generic RAG system

Memory Garden does retrieval — POST /v1/recall fuses lexical (FTS5/BM25), dense vector (sqlite-vec ANN), and graph-expansion signals — but the unit of retrieval is a typed atomic fact, not an opaque text chunk. Each embedding has an explicit (entity, relation, value) contract.

Not an agent runtime

Memory Garden sits above agent platforms (Claude Code, Paperclip, LangChain) and below the open internet. It does not orchestrate agents or run tools.

Not a compliance tool

The provenance trail Memory Garden produces is useful for audit — but building GRC workflows on top of it is the job of a separate application layer.

Not a general-purpose database

Facts are append-only knowledge records — typed, scoped, provenance-tagged, decay-aware. For arbitrary mutable state, use a real database.


Memory Garden / Stigmem — Apache 2.0. Spec · GitHub · Contributing