Skip to content

problem.md

Latest commit

 

History

History
64 lines (40 loc) · 2.5 KB

problem.md

File metadata and controls

64 lines (40 loc) · 2.5 KB

The Problem This Plugin Solves

This plugin exists because the stock OpenClaw memory path is optimized for lightweight persistence, not for a full context lifecycle with scope separation, compaction, and bounded retrieval.

What the Stock Plugin Is Good At

The default memory-lancedb style plugin is still a good fit when you want:

  • a simple top-k semantic lookup over durable notes
  • a low-complexity setup
  • a heuristic capture model where the agent decides what to save
  • no additional sidecar process

For short sessions and light persistent memory, that is often the right answer.

Where the Stock Model Breaks Down

The workload this plugin targets is different:

  • long technical sessions
  • repeated workflow state that must not be lost
  • user/session/global scopes that behave differently
  • compaction instead of unbounded append-only growth
  • retrieval that must fit within a prompt token budget

The failure modes are structural, not cosmetic.

Context collapse in long sessions

A single-table top-k memory system has no first-class notion of ephemeral session state versus durable user memory. A long active session accumulates turns that are highly relevant right now but should not pollute long-term recall forever.

No scope separation

This plugin uses distinct namespaces for:

  • session:<sessionId>
  • turns:<userId>
  • user:<userId>
  • global

Those scopes are not interchangeable. Session state needs different retention, different recency behavior, and different retrieval weighting than durable user memory.

No token budget management

Retrieval is only half the problem. The selected memories must still fit inside a bounded prompt budget. This plugin treats ranking and packing as one pipeline, not as disconnected concerns.

No automatic compaction

Raw turns are not the right long-term storage format. The plugin clusters session turns, summarizes them, removes source turns after confirmed summary insertion, and lets summary confidence affect later retrieval. Without compaction, memory becomes an append-only dump that eventually competes with itself.

Why This Is a Lifecycle Replacement

The solution is not a different LanceDB parameter or a larger topK.

The plugin replaces the whole context lifecycle:

  1. ingestion
  2. gating and promotion
  3. retrieval and ranking
  4. token-budget fitting
  5. compaction and summarization
  6. quality feedback into future retrieval

That is why the plugin is a slot-level memory replacement instead of a helper library layered beside the stock plugin.