TT-Wang/cortex-plugin

v0.11.0 Feature

This release adds 2 notable features for engineering teams evaluating rollout.

Published 1mo MCP Developer Tools

View tool

✓ No known CVEs patched

Read the diff → Tool health → What is this tool? →

✓ No known CVEs patched in this version

Topics

anthropic claude-code claude-code-plugin local-first mcp mcp-server

+4 more

memory obsidian persistent-memory python

Summary

AI summary

Session start memory injection defaults reduced and double-fire bug fixed, cutting token usage by ~5-6×.

Full changelog

Cuts per-session memem token cost by ~5-6× by adopting claude-mem's SessionStart defaults and adding project scoping.

Measured on the dev vault (1926 memories): SessionStart injection drops from ~20K tokens → ~3.7K tokens per session.

Changed

Defaults overhauled (claude-mem parity):

MEMEM_SESSION_START_LIMIT (default 50, range 1-200) — total memories injected. Previous default was 500.
MEMEM_SESSION_START_FULL (default 5, range 0-20) — of the top-ranked memories, how many show full content. Previous model loaded all L0 as full uncapped.
MEMEM_SESSION_START_PROJECT (default = cwd basename) — scope filter. memem now injects only memories from the current project's directory + cross-project general memories. Pass "all" to disable scoping.

Unified ranking across L0-L3: top-N by importance × recency. Layers still drive classification but no longer gate injection at session start. Claude gets a flat 'top N most useful for this project right now' set.

Fixed

Double-fire on first turn. SessionStart and UserPromptSubmit previously both fired on the first message, injecting duplicate material. SessionStart now writes ~/.memem/.last-brief.json with primed: true. The next UserPromptSubmit consumes the flag and returns empty context, so subsequent prompts use normal topic-shift logic.

Saves ~5-10K tokens on the first prompt of every session.

Estimated per-session savings

| Dimension | v0.10.2 | v0.11.0 | Savings |
|---|---|---|---|
| SessionStart memory count | 500 + up to 20 L0 full | 45 compact + 5 full | ~10× fewer |
| SessionStart tokens | ~20K | ~3.7K | ~16K/session |
| First-prompt double-fire | ~5-10K extra | 0 | ~7K/first turn |
| Scope | all projects | current project only | ~10× less noise |
| Total first-turn cost | ~25-30K tokens | ~3.7K tokens | ~7× |

Backward compat

Legacy MEMEM_COMPACT_INDEX_LIMIT still honored as a fallback
--limit N CLI argument still works
Memory layer classifier unchanged
Existing vaults load cleanly

Upgrading

/plugin update memem

Restart Claude Code to pick up the new SessionStart hook behavior.

View diff on GitHub

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Share on X Share on Bluesky

Track TT-Wang/cortex-plugin

Get notified when new releases ship.

About TT-Wang/cortex-plugin

Persistent, self-evolving memory plugin for Claude Code. Background miner extracts durable lessons (decisions, conventions, bug fixes)

All releases →