Changelog

What's new in each release

Version 0.13.0 — Per-agent context

July 2026

  • Per-agent standing context. Agents served by the same server can now each have their own context, in three modes: override (the agent's sections win per section, the rest falls back to global), global (shared set — still the default), or off. Zero config needed: drop Markdown files into ~/.hicortex/context/agents/<id>/ and that agent is overridden. Pin modes with contextAgents in config (wins over the drop-in dir).
  • Hermes and OpenClaw now receive the context layer. The Hermes plugin (0.6.0) and OC plugin inject a ## Context block above lessons at session start, scoped automatically per profile/agent (override with the optional agent_name setting). Context and lessons fail soft independently.
  • Claude Code stays global by design — your CC machines are all you. Opt a box into its own identity with hicortex init --agent-name <name>; clear with --agent-name ""; hicortex status shows it.
  • Edit any agent's context via the scope selector in /context/ui (inherited sections dimmed; saving one makes it an override) or hicortex context show|edit --agent <id>. Mixed fleets are safe: against an older server the new plugins hold off injecting rather than give every persona the same context.

Version 0.12.1 — Install fix

  • Fixed: npx client installs no longer break silently. When set up via npx @gamaze/hicortex init --server <url>, the session-start recall hook and the nightly capture job were wired to a temporary npx cache directory that npm later deletes — quietly stopping recall injection and capture. They now use the durable npx -y @gamaze/hicortex form. If you installed 0.12.0 as a client, re-run init (or upgrade to 0.12.1) to repair it.

Version 0.12.0 — Context layer

July 2026

  • Context layer — hand-edited standing context. Alongside auto-distilled memories and lessons, Hicortex now holds a hand-edited "who you are + how to work" layer — Markdown files (recommended starter sections user and rules, which you create) injected into every session at start. Unlike memories, it is never distilled, scored, or decayed: what you write stays verbatim until you change it. Edit it in the web editor at /context/ui or with hicortex context show|edit. Choose which harnesses receive it via contextClients in config (default ["cc"]; Claude Code only in v1). See Usage and Configuration.
  • Recall renamed context → recent (breaking). The queryless recall tool hicortex_context is now hicortex_recent, and REST GET /context is now GET /recent — same behavior, clearer name (the context name now belongs to the standing context layer above). The Hermes hicortex_recall_recent alias is dropped (merged into hicortex_recent). There is no compatibility alias: upgrade the server before its clients — a client on the old name loses queryless recall until it matches. Semantic hicortex_search is unchanged.

Version 0.11.0 — Memory visualization & self-organizing domains

July 2026

  • Memory graph visualization. The server now serves an interactive 3D knowledge graph of your agent's memories at /viz — domains as colored clusters, relationships as edges, hubs highlighted, with a one-click 2D toggle. Filter by domain, type, and strength; search to highlight; click any node to read the memory. Self-contained and served by your own server — no external requests.
  • Multi-tag memory domains. Define your own knowledge spheres in config — life areas or project/topic areas — and Hicortex files every memory into them automatically. Each memory gets every domain that applies, weighted from your own data, with one primary for the map. Memories that fit nothing fade away naturally — no "Unsorted" pile. See Domains.
  • Self-tuning classification. Domain profiles and weights are recomputed every night from the current data, so your categories reorganize as your work evolves — like human memory, nothing goes stale and nothing needs re-labeling.
  • Smarter memory links. Relationship discovery recalibrated — memories now link to genuinely related memories across the knowledge bank instead of near-duplicates only, making the graph (and /viz) far more connected. New hicortex relink command back-fills links across an existing memory bank.
  • init scaffolds default domains. New installs start with five editable domains (Work, Personal, People, Health, Finance). An existing domain list is never overwritten. New hicortex classify-domains command files a pre-existing memory bank into your domains in one resumable pass.

Version 0.10.0 — License change, all features free

July 2026

  • License changed to PolyForm Noncommercial 1.0.0. Personal and noncommercial use is free — the full product, no feature limits. Commercial use (for-profit businesses, client work, revenue-generating products) requires a per-seat license. See hicortex.gamaze.com for details. Versions ≤ 0.7.1 published to npm remain MIT-licensed.
  • All feature tiers and gates removed. No more 250-memory cap, no tier differences in functionality. Every install has the full feature set.
  • Per-install auth token. A unique token is generated at init and stored in your config. No more shared default token. Find it with hicortex status.
  • User-controlled LLM selection. init now presents a numbered list of detected LLM candidates and asks you to choose — nothing is silently auto-applied.

Version 0.9.0 — Thin-client capture

June 2026

  • New capture architecture: client machines no longer run local LLM distillation. Instead, the nightly job denoises sessions locally (no LLM — strips tool noise, truncates bulk I/O) and POSTs the denoised text to the server's POST /distill endpoint. The server handles all LLM work. Raw session content never leaves the originating machine.
  • New canonical endpoint: POST /distill accepts denoised session text and runs server-side distillation. Replaces the older pattern of client-side distillation + POST /ingest. The legacy /ingest endpoint is retained for backward compatibility.
  • distillFallback: "strict" (new default) — if the distill endpoint is unreachable, the nightly run aborts and retries next run instead of silently falling back to a lower-quality model.
  • Hermes support: nightly pipeline reads per-profile state.db alongside Claude Code sessions.

Version 0.8.0 — Hermes support

July 2026

  • Hermes recall plugin: recall-only adapter for Nous Hermes agents. Injects lessons into the system prompt at session start; exposes semantic search and queryless recent-memory recall tools in-process. Install via hermes memory setup hicortex.
  • Bedrock deployment: server runs on bedrock, lenovo connects as a client. Plugin v0.3.0 deployed to all profiles.

Version 0.7.0 — Graph Intelligence

April 2026

  • Graph-based community detection: Louvain algorithm discovers knowledge domains from the memory link graph — zero LLM cost, automatic during nightly consolidation.
  • Hub node detection: highly-connected memories get a strength boost, surfacing foundational insights in lesson selection.
  • New MCP tool: hicortex_graph enables graph traversal — find connected memories, hub nodes, or shortest paths between memories.

Version 0.6.0 — Knowledge Domain Routing

April 2026

  • Knowledge domain routing: memories are automatically grouped into knowledge domains during nightly consolidation. Agents see a structured domain index instead of a flat project list.
  • Domain-aware lesson selection: lessons from related projects in the same domain are boosted during injection.
  • New MCP tool: hicortex_index lets agents query what knowledge domains exist before searching.
  • Configurable injection token budget via moduleIndexTokenBudget in config.
  • Schema migration: new domain column on memories (automatic on upgrade).

Version 0.5.3 — Security & Quality

April 2026

  • Pre-ingestion redaction: API keys, tokens, paths, and secrets are automatically scrubbed from session transcripts before they reach the LLM or storage. 12 default patterns, configurable extras.
  • Contradiction detection: new lessons are checked against existing ones during nightly reflection. Contradicting lessons are suppressed to prevent wrong lessons from reinforcing themselves.
  • Anonymous telemetry for usage tracking (opt-out via config).
  • Pre-flight health check for reflect endpoint (skips reflection when reflect server is offline).

Version 0.5.0 — Pi Agent Support

April 2026

  • Pi coding agent support — nightly distillation reads Pi session transcripts alongside Claude Code sessions.
  • MCP tools work via pi-mcp-adapter connecting to the Hicortex SSE endpoint.
  • Package directory renamed from openclaw-plugin to hicortex.

Version 0.4.6 — Nightly Data-Loss Fix

April 2026

  • Fixed critical bug: nightly pipeline silently lost sessions when the distillation LLM was unreachable or the required model was missing.
  • Pre-flight health check on remote Ollama distill endpoints (reachability + model loaded).
  • Per-session deduplication in server mode — makes retries idempotent.
  • Last-run watermark only advances when every session was processed cleanly.

Version 0.4.5 — Public Release

April 2026

  • Hicortex client is now source-available at github.com/gamaze-labs/hicortex.
  • Pre-public hardening: scrubbed internal references, updated to current LLM model versions.

Version 0.4 — Multi-Client Support

March 2026

  • Connect multiple clients to one shared memory server.
  • Client mode: npx @gamaze/hicortex init --server <url>.
  • Split LLM configuration: separate models for scoring, distillation, and reflection.
  • New MCP tools: hicortex_update and hicortex_delete.
  • Balanced learning: reflection extracts lessons from both successes and failures.
  • Dynamic CLAUDE.md injection: lessons + memory index + project context.

Version 0.3 — Claude Code Support

March 2026

  • Same @gamaze/hicortex package works with both Claude Code and OpenClaw.
  • One-command install: npx @gamaze/hicortex init.
  • Persistent HTTP/SSE MCP server with 4 tools (search, context, ingest, lessons).
  • Automatic daemon install (launchd on macOS, systemd on Linux).
  • Nightly pipeline: auto-capture CC transcripts, distill, consolidate, inject lessons.
  • Claude CLI as LLM backend (uses your Claude subscription, no API key needed).
  • Custom commands: /learn, /hicortex-activate.

Version 0.2 — OpenClaw Plugin

February 2026

  • Initial release as OpenClaw lifecycle plugin.
  • Automatic session capture via OpenClaw hooks.
  • Nightly distillation and consolidation pipeline.
  • Semantic search with BM25 + vector embeddings + RRF fusion.
  • Knowledge graph with memory linking.
  • Memory decay and strengthening model.
  • Multi-provider LLM support (20+ providers).