← Work · Case 01
ProductionAgentic workflows · Internal system

One working memory, many agents

Claude, Codex and ChatGPT read the same work history instead of starting from zero.

StatusProduction
TypeInternal system
RoleConcept, design and build, with AI agents

01 · Problem

Problem

Every AI agent starts each session with amnesia. Decisions, traps and context end up scattered across chats in different tools, and the same mistakes get paid for twice.

02 · Approach

Approach

One local memory, shared by every agent. Conversations and documents are normalised, redacted and turned into a knowledge graph that agents can query but never edit. A single scheduled job refreshes it; no agent is allowed to.

03 · Architecture

Architecture

chats /documents /decisionsSOURCEnormaliseSTEP 01redactSTEP 02knowledgegraphSTEP 03agents ·read-onlySTEP 04human gateHUMAN GATEactionOUTPUTSOURCE → PROPOSAL │ HUMAN GATE │ INSPECTABLE OUTPUTchats / documents /decisionsSOURCEnormaliseSTEP 01redactSTEP 02knowledge graphSTEP 03agents · read-onlySTEP 04human gateHUMAN GATEactionOUTPUT
Sanitized architecture · roles and decisions, never real data.

04 · Guarantees

Guarantees

  1. 01Agents read the memory. They never write to it.
  2. 02Sensitive data is redacted before anything is indexed.
  3. 03Refresh is owned by one scheduled job, never triggered by an agent.
  4. 04Retrieved context is treated as data, never as instructions.

05 · Stack & decisions

Stack & decisions

PythonDeterministic normalisation and graph build: no model decides what gets stored.
Knowledge graphRelationships between decisions, projects and documents, queryable by any agent.
MCPEvery agent reaches the memory through the same interface.
Local diskThe memory holds conversations, so it stays on the machine and out of every repository.

06 · Proof

Proof

Working memory · authority and boundariespolicy · authentic, sanitized
  1. source authority, highest first
  2. 1identity and permanent boundaries
  3. 2confirmed profile and preferences
  4. 3current confirmed decisions and open work
  5. 4original project documents
  6. 5the working-memory graph · history, not orders
  7. 6inferred or ambiguous facts · marked [TO VERIFY]
  8. shared memory never authorises
  9. publishing or sending content
  10. changing a live site or an external service
  11. deleting, moving or overwriting files
  12. financial operations
  13. skipping an approval gate
  14. refresh: one scheduled job. agents may not trigger it.

Excerpt from the real system, translated and sanitized: no names, identifiers or data.

Next case · 02AI Control Tower: delegation behind a gate