SecondBrain/reference/memsearch-sessionstart-inje...

35 lines
1.6 KiB
Markdown
Raw Normal View History

2026-07-08 18:23:40 +00:00
---
summary: memsearch's SessionStart "Recent Memory" injection size is hardcoded in its hook script — no config knob controls it, and the documented TOML surface doesn't cover it.
tags:
- type/reference
- tool/memsearch
scope: global
type: reference
source: cc-os docs/plans/b3-memsearch-injection-review.md
date: 2026-07-08
last_updated: 2026-07-08
---
# memsearch SessionStart injection is hardcoded (no cap knob)
Verified on memsearch v0.4.6 and the v0.4.11 plugin cache (2026-07-08).
- The SessionStart `# Recent Memory` injection is produced by `hooks/session-start.sh`,
which hardcodes `head -2` daily memory files × `head -40` heading/bullet lines each.
- **No user-facing config knob controls this.** The documented TOML surface (the
`memory-config` skill: summarize / project_review / user_profile / memory_to_skill /
prompts) does not touch the injection size — don't waste time looking for a cap there.
- Measured cost across 61 real sessions (cc-os, 7 days): min ~6.0 KB / median ~11.7 KB /
max ~15.4 KB per session (~1.5k3.9k tokens), injected every SessionStart.
## Levers (in order of durability)
1. Upstream fix: a memsearch config option for injection size (doesn't exist yet; would be
the only durable cap).
2. Reduce bullet density at the source via the Stop-hook summarizer prompts/config — fewer,
denser bullets in the daily files shrink what `head -40` captures.
3. Unsupported stopgap: hand-edit the cached hook's `head -2`/`head -40` constants
(~75% reduction) — **any plugin cache refresh silently reverts it**.
Related: [[cc-os-hub]]