Reduce agent memory and MCP payload waste #124

Merged

kleb merged 4 commits from feature/memory-v2-efficiency into main

2026-08-30 04:06:02 +02:00

kleb commented

2026-08-30 03:51:46 +02:00

Collaborator

Summary

reject strict Codex title jobs and Claude task notifications before hook, session, storage, and recall work
replace growing hook responses and prompt-in-URL recall with compact compatible transports
compact high-volume MCP results and remove repeated server instructions from the MCP tool catalogue
inherit trusted session project scope and reject stopword-only relation candidates
align installed skill and MCP lifecycle tools around checkpoint-only summaries and one atomic final end

Measured motivation

machine notifications were 97.65% of stored Claude prompt text in the audited sessions
full hook session responses returned at least 8.52 MiB of redundant prompt JSON in two sessions
host-side repetition expanded 1,143 MCP instruction characters into 83,439 duplicate characters
ten task_list calls emitted 2.22 million result characters

Compatibility

existing stored sessions and prompts are unchanged
legacy full session responses and GET /memory/context remain supported
REST optimistic concurrency remains intact
no database schema or migration changes

Verification

bun run check:quality
953 server and unit tests passed
238 UI tests passed
targeted hook, MCP, prompt hygiene, scope, relation, and payload-bound tests passed

## Summary - reject strict Codex title jobs and Claude task notifications before hook, session, storage, and recall work - replace growing hook responses and prompt-in-URL recall with compact compatible transports - compact high-volume MCP results and remove repeated server instructions from the MCP tool catalogue - inherit trusted session project scope and reject stopword-only relation candidates - align installed skill and MCP lifecycle tools around checkpoint-only summaries and one atomic final end ## Measured motivation - machine notifications were 97.65% of stored Claude prompt text in the audited sessions - full hook session responses returned at least 8.52 MiB of redundant prompt JSON in two sessions - host-side repetition expanded 1,143 MCP instruction characters into 83,439 duplicate characters - ten task_list calls emitted 2.22 million result characters ## Compatibility - existing stored sessions and prompts are unchanged - legacy full session responses and GET /memory/context remain supported - REST optimistic concurrency remains intact - no database schema or migration changes ## Verification - bun run check:quality - 953 server and unit tests passed - 238 UI tests passed - targeted hook, MCP, prompt hygiene, scope, relation, and payload-bound tests passed

kleb added 1 commit

2026-08-30 03:51:46 +02:00

Reduce agent memory and MCP payload waste 4855da79e3

kAgent commented

2026-08-30 03:51:46 +02:00

Collaborator

kReview review

Verdict: no findings

No findings to address across the reviewed diff.

Reviewed by kReview at 397961fd47. This comment is conservative and based only on the PR diff, metadata, and supplied repository context.

Est. cost ~$7.16 total (4.0M in / 15.0k out) · this run ~$1.22 (541.2k in / 2.2k out) / gpt-5.6-sol + gpt-5.6-terra.

## kReview review **Verdict:** no findings No findings to address across the reviewed diff. _Reviewed by kReview at `397961fd47`. This comment is conservative and based only on the PR diff, metadata, and supplied repository context._ _Est. cost ~$7.16 total (4.0M in / 15.0k out) · this run ~$1.22 (541.2k in / 2.2k out) / gpt-5.6-sol + gpt-5.6-terra._

kleb added 1 commit

2026-08-30 03:55:46 +02:00

Allow read tokens to post memory context 9fd12080ed

kleb added 1 commit

2026-08-30 04:00:43 +02:00

Preserve write-only memory session controls 9a685cf7bb

kleb added 1 commit

2026-08-30 04:04:28 +02:00

Preserve MCP concurrency and task access metadata 397961fd47

kleb merged commit 1320de7c5d into main

2026-08-30 04:06:02 +02:00

kleb referenced this pull request from a commit