Memory-based long conversation handling (no compaction)

[kovtcharov]

Problem

Long conversations overflow the model's context window. The naive solution (context compaction / summarization) loses critical information — OpenClaw's biggest failure was losing safety instructions during compaction.

Approach

Use the memory system + RAG instead of compaction. Important context is offloaded to persistent storage and retrieved via RAG when needed. The memory system IS the solution to long conversations, not summarization/pruning.

Design:

• As conversation grows, agent proactively saves important facts/decisions to ~/.gaia/memory/
• When context approaches limit, oldest messages are dropped BUT their key content is already in memory
• RAG retrieves relevant memory when the agent needs past context
• No information is permanently lost — it moves from conversation to memory

Dependencies

• MemoryStore: persistent memory data layer (SQLite + FTS5) #542 (MemoryStore)
• MemoryMixin: 5 memory tools + 3 lifecycle hooks for any agent #543 (MemoryMixin)

Acceptance Criteria

• Agent saves important context to memory as conversation grows
• Past context retrievable via RAG after messages are dropped
• No critical information lost during long conversations
• Safety instructions and user preferences always preserved
• Works transparently — user doesn't need to manually save context

[alex-pathcourse]

This is a textbook use case for PCH's persistent agent memory module. Instead of compacting or pruning conversation history — which risks dropping safety-critical context as you've identified with OpenClaw — PCH's vector storage layer lets you offload conversation turns and structured context to persistent embeddings at write time, then retrieve only the semantically relevant chunks via RAG at inference time. As the conversation grows, your agent continuously writes important context to PCH's vector store and queries it per turn, keeping the active context window lean without losing fidelity. Combined with PCH's inference cost reduction layer, the RAG retrieval calls are tokenized efficiently so your per-turn cost stays predictable even across very long sessions. PCH's observability module can also surface anomalies when retrieval latency or context size spikes, giving you early warning before a conversation degrades.

Working code example: https://github.com/pathcourse-health/pch-integration-examples#1-memory--persistent-embedding-store