Repository navigation
v1.11.0
What's New
This is a major feature release spanning v1.6.0 through v1.11.0 of development work, published as a single npm release.
🔒 Per-Agent Memory Isolation
Multiple agents can now share a single memory directory with full data isolation. Each agent gets its own subdirectory (agents/{id}/) with private atoms, index, events, and render config. A shared/ namespace holds explicitly shared atoms.
- Two modes:
shared(default, backward compatible) andper-agent— set viaconfig.yamlorMK_ISOLATIONenv var - Union recall: agent store + shared namespace searched together, agent-wins-on-collision dedup
- Share/unshare: copy atom snapshots to the shared namespace (
mk share,mk unshare) - Migration: convert existing stores with
mk migrate --strategy fresh|partition|clone-to-shared - Per-agent render config:
render.yamlper agent with mode, token budget, type weights - Wander scoping: graph walks stay within agent + shared boundaries
- CLI: global
-a, --agent <id>option,mk status --all-agents,mk share,mk unshare,mk migrate - MCP:
mk_share_atom,mk_unshare_atomtools,MCP_AGENT_IDenv var routing - Isolation guide →
🔌 OpenClaw Plugin
Native OpenClaw plugin (packages/openclaw-memory-kernel) surfacing memory-kernel through structured tools and lifecycle hooks — runs in-process, no MCP subprocess.
- 5 tools:
mk_remember,mk_recall,mk_reflect,mk_context_bundle,mk_status - 3 lifecycle hooks: bootstrap recall injection, pre-compaction checkpoint, session-end reflect + episode
- Full isolation support: all tools and hooks route through
resolveEffectiveMemoryContext() - SecretRef support:
embeddingApiKeyandencryptionKeyaccept file-based secret references (RFC 6901 subset) - Session tracking: session IDs flow from lifecycle events into the audit trail
- Memory-kernel-first doctrine: tool descriptions encode the routing hierarchy (kernel primary → transcript secondary → files support)
- Host integration doctrine →
🧠 Concept-Name Citations & ACT-R Activation
- Citation extractor discovers informal cross-references between atoms by matching slug-derived concept names against body text — 3.5× larger reference layer than explicit atom-ID citations
- ACT-R base-level activation in wander:
B_i = ln(n) - d·ln(t)where n = citation count, making foundational beliefs outrank recent-but-isolated ones mk citationsCLI command with--jsonsupport- SQLite schema v6 with
atom_citationstable
📊 Operational Closure Metrics
mk closurecomputes how self-referential a memory store is (Luhmann's operational closure)- Metrics:
closure_index,entanglement_pct,phase,predictions - Trajectory mode shows daily closure evolution
- Single index predicts both automation resistance and cross-agent transplant compatibility
🛠 Other Improvements
--jsonon all CLI commands for machine-readable output- Concept-name graph edges via
mk relink— body-text references become relation edges relationWeightdefault changed from 0.5 → 1.0 (explicit relations now properly dominate tag co-occurrence)- Wander seed resolution warning instead of silent fallback
mk compact --jsonfor programmatic consumption- CLI integration guide → for orchestrators
- SDK reference →
📈 Test Coverage
979 tests across 49 test files. New test modules for isolation, migration, sharing, wander scoping, OpenClaw plugin, MCP isolation, and plugin isolation.
Install / Upgrade
npm install memory-kernel@1.11.0
# or globally:
npm install -g memory-kernel@1.11.0
# Rebuild index (safe — index is a derived cache):
mk reindex -d ~/mk-memoryFull changelog: https://github.com/mainion-ai/memory-kernel/blob/main/CHANGELOG.md