Skip to content

What's New

The latest releases — AutoMem server 0.16.0 and mcp-automem 0.15.0 (June 2026) — focus on recall correctness, batch operations, two new install paths, and data-driven tuning. This page summarizes the highlights; each item links to the page that documents it in full.

For the complete commit-level history, see the Changelog and the release notes on automem and mcp-automem.

AutoMem server 0.16.0

Section titled “AutoMem server 0.16.0”

Recall ranking overhaul

Section titled “Recall ranking overhaul”

The ranking release reworks how recall scores and orders results — the theme is correctness over knobs: fewer hand-tuned weights, more principled scoring.

  • Date-aware ranking & latest-fact selection — recent facts are prioritized while older context stays reachable, and superseded facts give way to the latest answer. See Recall Operations and Hybrid Search.
  • Configurable recency — the recency decay window and curve are now tunable, alongside an optional relative-recency re-rank. See the Recall Tuning guide.
  • Tag-score cap — the tag-score denominator can be capped to remove query-length bias (opt-in). See Recall Tuning.
  • Relevance gate — query-independent scoring is gated on topical evidence within the tag scope, cutting low-signal matches. See Recall Tuning.
  • state_mode=current|history — recall can now return only current memories or include superseded/invalidated ones for audit timelines. See Recall Operations.
  • Metadata sidecar search and metadata + updated_at surfaced in the detailed recall format.

Bulk memory associations

Section titled “Bulk memory associations”

A single API call can now create or update many memory relationships at once — far faster for batch graph building. See Relationship Operations.

Admin backup endpoint

Section titled “Admin backup endpoint”

A new admin endpoint exports your memories for backup and recovery. See Admin Operations and Backup & Recovery.

Consolidation: entity dedup & identity synthesis

Section titled “Consolidation: entity dedup & identity synthesis”

Consolidation gained entity deduplication and identity synthesis, plus hardened entity cleanup/repair tooling, so the graph collapses duplicate people/tools/entities more reliably. See Consolidation & Decay.

Enrichment metrics

Section titled “Enrichment metrics”

/enrichment/status now exposes the classification fallback-rate so you can see how often enrichment falls back from the LLM classifier. See Enrichment Pipeline and Health & Analytics.

Recall Quality Lab

Section titled “Recall Quality Lab”

A new contributor/maintainer harness tunes recall scoring against a clone of real data instead of guesswork — generate a test set with known answers, then measure Recall@K / MRR / NDCG with a statistical comparison to tell a real improvement from noise. It is not required to run AutoMem. See Hybrid Search and Research & Motivation.

mcp-automem 0.15.0

Section titled “mcp-automem 0.15.0”

Two new install paths

Section titled “Two new install paths”
  • Hermes Agent — AutoMem now installs into the Hermes terminal agent as MCP tools, a native memory provider, or both. See the Hermes guide.
  • Guided cloud installer — the installer can stand up a hosted backend (InstaPods or Railway) and capture the endpoint + token for you. See Guided Cloud Setup.

Plugin-first Claude Code

Section titled “Plugin-first Claude Code”

claude-code now installs as a Claude Code plugin that bundles the MCP server + hooks and auto-updates, replacing the manual settings-level wiring (still available as an option). Legacy hook entries are auto-migrated on re-run. See the Claude Code guide.

Smarter storage hooks

Section titled “Smarter storage hooks”

The Stop hook now uses an LLM-judged storage nudge instead of mechanically capturing build/test/deploy output — durable facts are stored by the model per the shared memory policy. See Memory Rules & Patterns.

Supersede memory mode

Section titled “Supersede memory mode”

store_memory can supersede an existing memory — store a correction and invalidate the old fact in one step. See Memory Operations.

Leaner recall responses

Section titled “Leaner recall responses”

Summary-first, token-aware budgeted recall formats keep responses well under the response-size cap so recalls don’t flood an agent’s context. See Recall Operations.

Upgrade notes

Section titled “Upgrade notes”
ChangeImpactWhat to do
AUTOMEM_ENDPOINT → AUTOMEM_API_URLThe endpoint variable was renamed (the old name still works as a fallback)Prefer AUTOMEM_API_URL in new configs. See Environment Variables.
Installer permission scope tightenedThe installer now merges only the six mcp__memory__* entries instead of granting broad Bash/file-tool permissions; re-running removes the four hook-era grantsRe-run the installer to adopt the narrower permission set
Hooks no longer auto-capture resultsBuild/test/deploy output is no longer mechanically stored; the model stores durable facts insteadNo action required; expect fewer, higher-signal memories

Next steps

Section titled “Next steps”