Why
Per-query synthesis already ships in recall (v1.4.0, issue #59) — that solved the "dumb retrieval" problem for single queries. What's still missing is the write-back loop: when a tag accumulates dozens of raw entries, there's no compression happening. Users returning to a heavily-used tag get 15 raw memories instead of one clean synthesis. The retrieval quality problem is solved ephemerally; the storage bloat problem is not.
What this covers
- Rolling summaries: a nightly cron job identifies tags with >20 entries and compresses related clusters into a single
synthesized entry, preserving source IDs in metadata as provenance links. Originals are not deleted.
/digest?tag=X endpoint: on-demand synthesis of everything stored under a given tag, returning a structured "state of the world" paragraph rather than raw entries. Useful for project handoffs, onboarding a new AI session, or sharing project context.synthesized tag convention: summary entries are tagged synthesized and linked back to source IDs so callers can distinguish generated summaries from raw memories.
Out of scope (already shipped)
Per-query synthesis on recall responses (issue #59, v1.4.0). This issue is about writing synthesized entries back to storage, not enriching retrieval responses.
Implementation notes
Why
Per-query synthesis already ships in
recall(v1.4.0, issue #59) — that solved the "dumb retrieval" problem for single queries. What's still missing is the write-back loop: when a tag accumulates dozens of raw entries, there's no compression happening. Users returning to a heavily-used tag get 15 raw memories instead of one clean synthesis. The retrieval quality problem is solved ephemerally; the storage bloat problem is not.What this covers
synthesizedentry, preserving source IDs in metadata as provenance links. Originals are not deleted./digest?tag=Xendpoint: on-demand synthesis of everything stored under a given tag, returning a structured "state of the world" paragraph rather than raw entries. Useful for project handoffs, onboarding a new AI session, or sharing project context.synthesizedtag convention: summary entries are taggedsynthesizedand linked back to source IDs so callers can distinguish generated summaries from raw memories.Out of scope (already shipped)
Per-query synthesis on
recallresponses (issue #59, v1.4.0). This issue is about writing synthesized entries back to storage, not enriching retrieval responses.Implementation notes
/chatendpoint pattern (Workers AI + Llama already in place)deleteByIdsis not called on originals — summary is additive