Context

Implement the shared organizational memory system defined in DESIGN_SPEC §7.4. Beyond individual agent memory (§7.1-7.3, covered by #32), the framework needs organizational memory — company-wide knowledge that all agents can access: policies, conventions, ADRs, coding standards, and operational procedures.

This is not personal episodic memory ("what I did last Tuesday") but institutional knowledge ("we always use FastAPI, not Flask").

Shared organizational memory is implemented behind an OrgMemoryBackend protocol, making the system modular and extensible.

Backends (§7.4)

Backend 1: Hybrid Prompt + Retrieval (Default / MVP)

Critical rules (5-10 items) injected into every agent's system prompt. Extended knowledge (ADRs, procedures) stored in a queryable store, retrieved on demand at task start.

• Simple to implement. Core rules always present. Extended knowledge scales.
• Basic retrieval may miss relational connections.

Backend 2: GraphRAG Knowledge Graph (Future)

Organizational knowledge as entities + relationships in a knowledge graph. Multi-hop reasoning: "FastAPI is our standard" → "don't use Flask" → "exception: data team uses Django for admin."

• Significant accuracy improvement over vector-only retrieval.
• More complex infrastructure.

Backend 3: Temporal Knowledge Graph (Future)

Like GraphRAG but tracks how facts change over time. Agents see current truth but can query history.

• Handles policy evolution naturally.
• Most complex. Potentially overkill for small setups.

Acceptance Criteria

Protocol Interface

• OrgMemoryBackend protocol defined: query(context) → list[OrgFact], write(fact, author), list_policies()
• Backend selection configurable via YAML (org_memory.backend: "hybrid_prompt_retrieval")
• New backends addable without modifying existing ones

Hybrid Prompt + Retrieval Implementation (MVP)

• core_policies list injected into every agent's system prompt
• Extended knowledge store with queryable interface (backed by SQLite initially)
• Retrieval at task start with configurable max_retrieved_per_query
• OrgFact model: content, author, category, created_at, version

Write Access Control

• Core policies: human-only writes
• ADRs and procedures: writable by senior+ agents (configurable role levels)
• All writes versioned and auditable
• Write access configurable per fact category

Testing

• Unit tests for protocol and MVP backend (>80% coverage)
• Test: core policies always present in system prompt
• Test: extended knowledge retrieved by relevance
• Test: write access control enforced per role level

Dependencies

• Individual agent memory interface (Design individual agent memory interface: working, episodic, semantic, procedural (DESIGN_SPEC §7.1-7.3) #32) — separate but related
• Memory layer candidate evaluation (Evaluate memory layer candidates: Mem0, Zep, Letta, Cognee, custom #39) — selected library may provide graph capabilities for Backends 2-3
• System prompt construction (engine, Implement agent engine core with ExecutionLoop protocol integration (DESIGN_SPEC §3.1, §6.1, §6.5) #11) — for injecting core policies

Design Spec Reference

• §7.4 — Shared Organizational Memory (OrgMemoryBackend protocol, 3 backends, write access control)