Further Reading
Background and related reading on agent memory, MCP, and the problem space Open Brain operates in.
Agent memory
- Agent Memory: Why Your AI Has Amnesia and How to Fix It — Casius Lee (Oracle Developers Blog, Feb 2026). Breaks down the four memory types (working, procedural, semantic, episodic), why bigger context windows aren't a substitute for persistent memory, and why agent memory is fundamentally a database problem.
Anthropic engineering on agentic harnesses
- Harness design for long-running application development — Anthropic Engineering (April 2026). Multi-agent (Planner / Generator / Evaluator) patterns, context resets vs. compaction, and the principle that harnesses encode assumptions that go stale as models improve.
- Claude Managed Agents — Anthropic Engineering (April 2026). Architecture of a durable event-log session, stateless harness processes, and stable interfaces (
execute,getEvents,wake) that decouple the session/harness/sandbox trio.
Protocol
- Model Context Protocol — The open standard every Open Brain tool speaks. Any MCP-compatible client (Claude Code, Cursor, Windsurf, ChatGPT Desktop) can connect.
brain_v2 design sources
The v2 redesign was informed by parallel synthesis documents from three AI agents, each independently analyzing v1's failure modes and proposing architectural fixes:
- Windsurf memory-architecture synthesis (internal design note) — selected as best-of-breed by Dave. Key contributions: headline-only boot with token cap, typed stores, ephemeral working context, temporal+spatial locality retrieval.
- Claude memory-architecture synthesis (internal design note) — complementary analysis. Key contributions: write gate atomicity checks, immutable rule bodies.
- GPT memory-architecture synthesis (internal design note) — complementary analysis. Key contributions: predictive context positioning, composite scoring.
- Infra-cost addendum (internal design note) — falsifiable Ollama runtime cost check. Proved v2's write path does not trigger metadata LLM eviction.