Repository navigation
B1: Symbolic drawer index over verbatim + micro-benchmark gate (flagged) #12555
Description
Part of #12551 (Gap B). Conditional — merged flag-on only if the benchmark wins.
Add a MemPalace-style symbolic pre-index to memory/verbatim_store.py so narrow-scope recall can skip a full vector scan.
Scope
- On
VerbatimStore.append: extract salient entities/keywords and write an inverted entity/keyword → chunk_id index + a coarse 'drawer' grouping (session/entity) in Redis ('knowledge' DB). - New optional search path: resolve candidate drawers symbolically, then vector-rerank within them (fall back to current full path on miss).
- Behind flag
AUTOBOT_VERBATIM_SYMBOLIC_INDEX(default off). - Ship a micro-benchmark (fixed corpus) comparing latency + recall@k: symbolic-prefilter vs current path.
Gate (hidden-cost)
Merge flag-on ONLY if recall-neutral (no regression) AND latency improves. Otherwise land flag-off and record the benchmark numbers on this issue as the rationale.
Acceptance
- Index stays consistent across append +
delete_session. - Benchmark posted with numbers.
- Flag-off path byte-identical to today.
Activity
fix(memory): cap symbolic candidates, recency parity, honest benchmark (
feat(memory): symbolic drawer index for verbatim recall, benchmark-ga…
fix(memory): cap symbolic candidates, recency parity, honest benchmark (
feat(memory): symbolic drawer index for verbatim recall, benchmark-ga…
fix(memory): cap symbolic candidates, recency parity, honest benchmark (
feat(memory): symbolic drawer index for verbatim recall (benchmark-ga…
mrveiss commentedon Jul 26, 2026
Delivered in PR #12580 (merged to Dev_new_gui). Ships flag-off (AUTOBOT_VERBATIM_SYMBOLIC_INDEX); flip on only after the benchmark proves a latency win with no recall regression. Fast-follows: #12581.
feat(agent): spill and anchor oversized tool output (#13692 step 1)
feat(agent): spill and anchor oversized tool output (#13692 step 1) (#…