thedotmack/claude-mem v15.18.0

on GitHub

eve agents get the same memory as Claude Code and Codex

claude-mem/eve now gives an eve agent the full claude-mem memory loop through your cmem.ai account (#6042).

• Session context: each turn starts with the recent timeline from your memory. It is refreshed every turn under one message id, so it replaces itself instead of piling up.
• Recall each turn: one search before the turn, as before.
• Recording is now on by default. After each completed turn the provider sends the prompt, one observation per tool call, and a turn summary to cmem.ai as platform eve. Recorded tool calls and summaries come out of your cmem.ai observer allowance, the same allowance your other clients use. Pass capture: false to keep the old read-only behaviour; session context, recall, search and remember keep working.
  • <private>…</private> regions and secret-shaped text are removed on your deployment before anything is sent, with the same rules the Cowork plugin uses.
  • Replays of the same turn are dropped by cmem.ai (turn-scoped tool call ids and an idempotency key).
  • Each request is sent once. No retry, spool or queue, so an outage never creates spend later. A failure is logged with how many events were accepted.
• Tools: search, timeline, get_observations (the three-step Smart search) and remember.
• eve versions: tested on eve 0.84.0 and 1.0.0-rc.9.
• Template: works on a free-credit Vercel AI Gateway (Claude, falling back to openai/gpt-5-mini when Claude is refused on free credits), or set ANTHROPIC_API_KEY to call Claude directly and EVE_MODEL to pick an exact model. Scheduled runs get the owner's memory, owner email must match exactly, and a morning-digest example schedule ships turned off.

Docs: https://docs.claude-mem.ai/eve-integration

Install choices move to the browser

A standard install now sets up the runtime in the terminal, then opens the local web setup to choose agents and integrations, bundled skills, model, native-memory preferences, optional email updates and telemetry (#5967). Agent shells can pass --setup browser when a person is there, or --setup defer to leave the choices for the first My Work visit. Explicit --provider / --ide installs and npx claude-mem update behave as before.

Also new

• Gemini CLI: native lifecycle capture with npx claude-mem gemini install|status|uninstall (#5991, thanks @Kirat0201, closes #5949). Docs: https://docs.claude-mem.ai/gemini-cli/setup
• Knowledge corpus on your API provider (opt-in): CLAUDE_MEM_CORPUS_AGENT=api answers query_corpus through the selected OpenRouter or OpenAI-compatible provider instead of the Claude Agent SDK. The default is unchanged (#5682, thanks @bulanovk).
• Agent SDK: a project-scoped memory example (#5995, thanks @Kirat0201).

Privacy clarification

• Prompt-history uploads (#5176): browser setup asks before sending selected prompt history to cmem.ai (fixed in v15.8.3, #5220). Skip this step sends no prompt history. The owner-goals skill can use your existing coding agent to analyze local history and save through the local worker without a separate cmem.ai intent upload. Inference still uses that agent's model provider. Updating does not remove data previously sent.

Fixes

Thanks @Afloat16 for every fix below unless another name is given.

Search and memory

• Unicode words in long questions still match indexed memories (#5997); canonically equivalent titles match (#5975); literal backslashes in POSIX file history are kept (#5985, thanks @Kirat0201); the observation-only semantic endpoint paginates (#5892).
• Recall survives repeated project merges (#5990), and Chroma keeps live metadata during a merge and bootstraps large projects without spreading row ids (#5881, #5862).
• Distinct structured observation content is no longer merged (#5979); colliding manual session identities are kept apart (#5868).
• Native memory notes: quoted metadata is decoded (#5983), BOM-prefixed Markdown parses (#5976), and frontmatter needs a real top-level closing line (#5845).
• Knowledge answers in several parts are kept whole (#6002).

Context and server

• Server context rows obey mode filters (#5993, thanks @0xABCD01) and keep their platforms when prior replies are selected (#6003); cached labels refresh after discovery usage settles (#6006).
• The server keeps platform scope for unlinked manual memories (#5875), native receipt identity through event validation (#5872), event platforms in viewer feeds (#5846), and capture mutation receipts within the key's scope (#5883).
• Delegated recall keys need read scope (#5856); in-flight generation results are purged before returning (#5858); salvaged summary bodies show in the viewer (#5869).
• Shutdown stops accepting HTTP before draining connections (#5853).
• The observer keeps delivery identity when capture is disabled (#5867); sync keeps the outbox when a stopped push is rejected (#5850).

Import and setup

• Imports keep compressed Codex history (#6000), Gemini text-part provenance and JSONL history edits (#5999, #5865), Unicode in prompt provenance (#5989), distinct legacy Claude prompts (#5982) and legacy observation text (#5854); copied OpenCode fork activity is excluded (#6004).
• A history import is bounded to its starting rows (#5877); local scans include archived Codex history (#5888); token coverage leaves out observer history and Codex context estimates (#5847, #5886).
• Exports include the full saved history (#5879); linked data-directory ancestors work for media (#5874); Claude Code rule imports escape spaces (#5852).
• Telemetry discloses default sharing in the first session (#5884).

My Work

• Shared-work sync passes are bounded so a hung thread cannot stall owner actions (#6014, thanks @LuckVd), and reader requests have a deadline (#5860, thanks @LuckVd).
• Session provenance is republished after a failed activity publish (#6010, thanks @bulanovk); failed shared-activity publication is retried (#5870).
• Owner goals drop unverifiable quotes from a fresh save instead of refusing the whole save (#5992, thanks @0xABCD01); rules-only goal snapshots are reused (#5987); newer focus decisions survive an undo (#5849).
• Products refresh when local files change (#5887), GitHub products link from SSH push output (#5998), valued git push options parse before targets (#5885), and single-quoted commit subjects are kept literally (#5896).
• Deliverables keep declarations for previously captured copies (#5973).

Integrations

• OpenCode: the current V2 built-in tool contract is normalized and failed tool status is kept (#5857, #5855).
• OMP: native tool semantics and tool-call receipt identity are kept, and hooks install into the selected profile (#5871, #5864, #5876).
• OpenClaw: JSON5 configuration is accepted, and the selected host config and state paths are used (#5861, #5844).
• Windsurf: user hook commands survive installation and the MCP server identity is kept for recursion filtering (#5882, #5880).
• Cursor: MCP argument and result JSON envelopes are decoded (#5848). DSH: the managed transcript path refreshes on reinstall (#5873). T3 Code: TOML plugin state is read before removal (#5851).
• Smart file reading includes C and C++ header declarations, compact Java record constructors and TypeScript declare functions (#6015, #6012, #6005); XML literals keep their closing tags (#5977).

Redaction and performance

• Pulumi access tokens and Doppler personal tokens are redacted (#5984, thanks @kevin9327).
• Telemetry aggregates event statistics on arrival (#5994), long transcript lines are assembled once on import (#5981), OpenCode token usage is streamed (#5974), and legacy prompt retry lookups in the viewer are bounded by time (#6011).

Docs

• Agent framework guides are grouped, and a setting that does not exist was removed (#6030).
• The token claim is stated as a per-result size ratio (#5889, thanks @Zhong-fan, #4138).

Full changelog: v15.17.1...v15.18.0