Context

The memory layer library is TBD per DESIGN_SPEC §15.2. Multiple candidates are under evaluation. Before committing to any library, we need a thorough comparative evaluation to determine the best fit for our requirements.

Candidates under evaluation:

• Mem0 — Universal memory layer for AI agents (open-source, production-proven)
• Graphiti — Temporal knowledge graph for AI agents (Zep's open-source successor)
• Letta — Stateful LLM agents with memory management (formerly MemGPT)
• Cognee — Knowledge engine for AI agents (graph-based)
• Custom solution — Neo4j + Qdrant + FastEmbed/LiteLLM composed stack
• Plus 12+ additional candidates discovered during open search

Evaluation Criteria

For each candidate, assess:

• Does it support all our memory types (working, episodic, semantic, procedural, social)?
• Retrieval quality — published benchmarks and community reputation
• Per-agent memory isolation support
• Fully local operation (no cloud dependencies required)
• Memory consolidation/summarization capabilities
• API compatibility with our abstract memory interface (§7.1-7.3)
• Compatibility with OrgMemoryBackend protocol (§7.4) — does the library provide graph memory capabilities that could serve as a backend for organizational memory?
• Active maintenance, community size, license compatibility (must be compatible with BUSL-1.1)
• Python 3.14+ compatibility
• Async support quality
• Container architecture and operational complexity
• Embedding provider configurability (local + cloud)

Acceptance Criteria

• Open discovery: broad search beyond spec-listed candidates (16+ found)
• Gate checks: pass/fail on local-first, license, Docker, Python 3.14, isolation, embeddings
• Comparison table: all gate-passing candidates scored against weighted criteria (S1-S11, 100 points)
• Decision documented with rationale: ADR-001 in docs/decisions/ADR-001-memory-layer.md
• Evaluate whether selected candidate can serve §7.4 Shared Organizational Memory backends
• Downstream impact documented for Design individual agent memory interface: working, episodic, semantic, procedural (DESIGN_SPEC §7.1-7.3) #32 (memory interface), Implement pluggable PersistenceBackend protocol with SQLite backend (DESIGN_SPEC §7.5) #36 (persistence), Implement shared organizational memory with OrgMemoryBackend protocol (DESIGN_SPEC §7.4) #125 (org memory)
• DESIGN_SPEC.md §15.2 updated to reflect decision
• Backend swappability strategy: protocol-based design allows future backend swaps via config

Decision Summary

Mem0 as initial backend (in-process, Qdrant embedded + SQLite, persistent to Docker volume) behind pluggable MemoryBackend protocol. Custom stack (Neo4j + Qdrant external) as planned future upgrade. Cognee/Letta on watch list pending Python 3.14 support.

Key findings:

• 16+ candidates evaluated, 3 passed all gates (Mem0, Graphiti, Custom Stack)
• Letta and Cognee eliminated by Python <3.14 constraint (conservative bounds, not technical — on watch list)
• Supermemory eliminated: proprietary engine, SDK-only open source
• Mem0 chosen as initial for production-readiness (v1.0+, 49k stars) and low setup cost
• Protocol architecture ensures any backend can be swapped in later via config

Dependencies

• Depends on memory system interface (Design individual agent memory interface: working, episodic, semantic, procedural (DESIGN_SPEC §7.1-7.3) #32)

Design Spec Reference

• §7 — Memory System (§7.1-7.3 individual agent memory, §7.4 shared org memory)
• §15.2 — Technology Stack (updated with decision)

Updated 2026-03-08: Evaluation complete. ADR-001 accepted. Acceptance criteria revised to match research-based evaluation approach (prototypes and latency benchmarks deferred — unnecessary given pragmatic decision to use Mem0 as initial backend with protocol-based swappability).