Eidetic Labs · Infrastructure for agent work
Eidetic — recalled with photographic clarity. We build two pieces of the agent stack: Stigmem, a federated memory fabric, and Craik, a governed agent runtime. Each stands alone. Together, they compose.
Most agent platforms invest in tool calls, prompt routing, parallel runs.
Those are necessary. They are not enough.
An agent is useful at organizational scale only when it can remember, justify, coordinate, dispute, and hand off its work — across sessions, tools, runners, and the humans who review it. Eidetic Labs builds that substrate as two composable pieces: Stigmem for the memory, Craik for the run.
02 Stigmem
A shared memory layer for AI agents. Typed, provenance-tagged facts that travel across tools, platforms, and organizations.
$
pip install stigmem
preview alpha · v0.9.0a2 · Apache-2.0
03 Craik
Shared project state, policy-bound authority, durable receipts, and handoffs the next agent can trust. Runner-neutral: Codex, Claude, Gemini, direct OpenAI & Anthropic, OAI-compatible locals.
$
pip install craik
governed MVP · 0.1.x · MIT
04 Composition
Stigmem holds the truth. Craik governs the act. The interface is small and explicit: Craik reads facts as evidence, writes receipts back as facts, and surfaces contradictions to the inbox they share.
$ stigmem assert migration_0042.row_count = 12M
→ fact_8a · signed · scope=team-dba
$ craik case build task_042
→ pulls fact_8a · 12 evidence refs
$ craik run task_042 --policy strict.write
→ contradiction: stigmem fact_91 differs
→ surfaced to reviewer:@dba
Adopt Stigmem alone for cross-platform memory. Adopt Craik alone in its degraded local mode. Adopt both for the full loop — the seams are documented, the contracts are typed.
05 Principles
A short list, kept honest by the work. We update it when the work proves us wrong.
01
If an agent's recall is a side effect of its prompt, it isn't recall — it's improvisation. Memory belongs in the runtime, with structure, scope, and contradiction.
02
Every consequential action — provider call, write, credential use, policy decision — leaves a durable, structured record. Retries without receipts are just louder mistakes.
03
Write authority, review gates, and approval surfaces are runtime objects, not prompt suggestions. Prompts persuade; policy decides.
04
Codex, Claude, Gemini, locals — same contracts, same handoffs. The governance model outlives the model release cycle.
05
A run that doesn't produce a handoff the next agent can use is a run that didn't finish. We design backwards from the handoff.
06
We publish what we ship, what we know, and what we got wrong. Trust in agent work is earned by showing the work.
06 The lab
Eidetic Labs is an independent research-and-engineering studio. We prototype, ship, write up the result, and keep going. The aim is not a platform-shaped product surface — it is a substrate that holds up to real software work.