Skip to content

v0.2.0 — multi-platform + curator loop

Choose a tag to compare

View all tags
@TJCurnutte TJCurnutte released this 25 Apr 03:35
· 60 commits to main since this release
v0.2.0
68b07ad

Memory Hive v0.2.0 — multi-platform + curator loop

This release closes the curator loop and broadens platform reach. Every agent
contribution can now be searched, clustered, validated, and promoted with a
single command — and Memory Hive auto-wires into 22+ agent platforms beyond
Claude Code.

Highlights

  • 11 new CLI verbs for observability and curator workflow
  • 22 agent platforms auto-wired (Claude Code, OpenClaw, Hermes, Cursor, Continue, Gemini CLI, Goose, Amazon Q, OpenHands, Roo, Kilo, Windsurf, Warp, Sourcegraph Amp, Codex CLI, OpenCode, Aider, Open Interpreter, Cline, Zed, Crush) with idempotent managed-block injection
  • memory-hive doctor now multi-platform; surfaces stale raw learnings
  • CI smoke tests on Ubuntu + macOS for every new verb

Install / upgrade

curl -fsSL memoryhive.neural-forge.io/install.sh | sh

Existing installs: re-running the installer is idempotent — silos are
preserved, shared-hive files drop alongside as .upstream for diff.

Added — Curator + observability verbs (11 new commands)

A wave of parallel feature work added eleven new CLI verbs that close the
loop between agent contributions and curator promotion. All are pure shell,
all ship a CI smoke test, all read from the existing two-layer architecture
without changing it.

  • memory-hive tail [-n N] [--silo <name>] [--since <YYYY-MM-DD>] —
    Like tail -f for the hive (non-streaming). Prints the N most recent
    hive writes with key content extracted per kind: log/DECISIONS/CONFLICTS
    show the newest dated heading + first paragraph; memory/context show the
    tail; learnings show H1 + frontmatter. Use watch for the streaming view.

  • memory-hive promote <raw-file> [--into <name>] [--title <text>] [--dry-run] —
    One-command curator workflow. Reads a raw learning, appends a summary
    with backlink to a distilled file, and logs the decision in
    curator/DECISIONS.md. Raw file stays untouched as source of truth.
    --into defaults to inferring from frontmatter kind:.

  • memory-hive confidence — Clusters raw learnings by normalized title
    and suggests upgrades when a cluster crosses an architecture threshold
    (3 aligned low → medium, 3 aligned medium → high). Suggest-only; curator
    decides what to promote.

  • memory-hive tag <file> <tag1> [tag2 ...] + memory-hive tags [--agent <name>] [--tag <name>] —
    Lightweight topical tagging for learning files. tag adds to YAML
    frontmatter (lowercase, dashes, 2-24 chars, sorted alphabetically,
    deduplicated). tags lists every tag in the hive with file count;
    --tag <name> lists files using a tag; --agent <name> limits the scan
    to one silo. Lets topical clusters emerge so the curator can spin up new
    learnings/distilled/<topic>.md files when a tag passes a threshold.

  • memory-hive stale [--days N] [--agent <name>] [--count] + doctor
    integration — Surfaces raw learnings >N days old (default 7) with no
    curator decision. doctor lists top 5 inline; auto-escalates to a
    warning when count exceeds MEMORY_HIVE_STALE_THRESHOLD (default 20).
    Files already referenced in DECISIONS.md are excluded.

  • memory-hive checkpoint [--name <name>] | --list + memory-hive diff [--since <checkpoint|YYYY-MM-DD>] [--verbose] —
    Save a named reference marker, then show every hive write since that
    point. Default --since is the most recent checkpoint, or the install
    baseline if none exists. Lists New vs Modified files, classified by
    agent. Honors .baseline-installed so installer scaffolding never
    appears in diffs.

  • memory-hive dedup [--per-agent] [--strict] [--threshold N] —
    Cluster near-duplicate raw learnings by normalized title. Default mode
    is cross-agent (catches when two agents independently document the same
    finding); --per-agent restricts within a single silo.

  • memory-hive query <term> [--silo <name>] [--kind <k>] [--since <date>] —
    Searchable hive. Greps every text surface (silo log/memory/context,
    learnings raw + distilled, knowledge, registry, curator) with
    line-number context. Filter by silo, kind, or date.

  • memory-hive citations — Cross-agent citation graph. Walks every
    file for [..](path) links and builds a who-cites-whom map. Surfaces
    team topology without manual tracking.

  • memory-hive lint [--fix] [--strict] [<path>] — Frontmatter schema
    validation for learning files. Checks required keys, date format, valid
    agent, valid confidence. --fix rewrites where safe; --strict exits
    non-zero on warnings.

  • memory-hive digest [--today | --yesterday | --week | --since YYYY-MM-DD] —
    Human-readable change summary for a time window (default rolling 24h).
    Per-agent newest writes with extracted content, ordered newest-first.
    "What's been happening?" without manual grep.

  • Multi-platform integration. The installer now auto-detects and wires up
    every major agent platform that has a stable plain-text config file on
    local disk. 17 auto-inject targets, 5 manual-instructions platforms, plus
    an opt-in GitHub Copilot repo wiring:

    Platform Integration Config file
    Claude Code auto-inject ~/.claude/CLAUDE.md
    OpenClaw auto-inject ~/.openclaw/CLAUDE.md
    NanoClaw auto-inject ~/.config/nanoclaw/AGENTS.md
    Hermes Agent auto-inject ~/.hermes/memories/MEMORY.md
    Cursor auto-inject ~/.cursor/rules/memory-hive.mdc
    Continue.dev auto-inject ~/.continue/rules/memory-hive.md
    Gemini CLI auto-inject ~/.gemini/GEMINI.md
    Goose (Block) auto-inject ~/.goosehints
    Amazon Q Developer CLI auto-inject ~/.aws/amazonq/rules/memory-hive.md
    OpenHands auto-inject ~/.openhands/microagents/memory-hive.md
    Roo Code auto-inject ~/.roo/rules/memory-hive.md
    Kilo Code auto-inject ~/.kilocode/rules/memory-hive.md
    Windsurf (Codeium) auto-inject ~/.codeium/windsurf/memories/global_rules.md
    Warp auto-inject ~/.agents/AGENTS.md
    Sourcegraph Amp auto-inject ~/.config/amp/AGENTS.md
    OpenAI Codex CLI auto-inject ~/.codex/AGENTS.md
    OpenCode auto-inject ~/.config/opencode/AGENTS.md
    Aider manual snippet ~/.aider.conf.yml (structured YAML)
    Open Interpreter manual snippet profile YAML
    Cline (VS Code) manual snippet VS Code settings UI
    Zed manual snippet ~/.config/zed/settings.json
    Crush (Charm) manual snippet ~/.local/share/crush/crush.json
    GitHub Copilot (repo) opt-in (MEMORY_HIVE_COPILOT_REPO=1) $PWD/.github/copilot-instructions.md

    Every auto-inject target uses the same <!-- memory-hive:start --> /
    <!-- memory-hive:end --> markers, so re-installs are idempotent and
    user content outside the markers is always preserved.

  • Per-platform opt-out env vars. Every platform has a matching
    MEMORY_HIVE_SKIP_<PLATFORM>=1 env var (e.g. MEMORY_HIVE_SKIP_CURSOR=1,
    MEMORY_HIVE_SKIP_GOOSE=1). The legacy MEMORY_HIVE_SKIP_CLAUDE_MD=1
    still works as an alias for MEMORY_HIVE_SKIP_CLAUDE_CODE=1.

  • Platform docs under templates/platforms/.
    One markdown file per supported platform explaining what gets wired,
    where, and how to opt out. Shipped into ~/.memory-hive/templates/platforms/
    on install so users can read them without cloning the repo.

  • memory-hive doctor is now multi-platform. Instead of only checking
    ~/.claude/CLAUDE.md, doctor iterates every auto-inject platform whose
    config file exists on disk and reports per-platform whether the managed
    block is present and points at the right install directory.

  • New CI smoke tests for the platform loop: detection, auto-inject,
    manual, opt-out, idempotency, and the multi-platform doctor. Runs on
    both ubuntu-latest and macos-latest.

Added (prior Unreleased work)

  • Three-tier memory flow documented in HIVE_ARCHITECTURE.md:
    raw external capture (Tier 1) → agent synthesis (Tier 2) → curator
    distilled truth (Tier 3). Clarifies how external context enters the
    hive without changing any existing flow.
  • hive/raw/ folder convention for Tier 1 raw external capture.
    One subfolder per source (hive/raw/discord/, hive/raw/slack/,
    etc.). Append-only; never modified by agents. Includes a README.md
    explaining the convention.
  • templates/memory-entry.md — canonical format spec for Tier 1
    raw capture and Tier 2 structured synthesis. Two shapes, one spec.
  • examples/ingesters/discord/ — reference Discord ingester.
    Polls channels via the bot API and appends to hive/raw/discord/.
    Pure stdlib, env-var configured, resumable via state file,
    append-only. Includes a README with setup, launchd/cron recipes,
    and adaptation notes.
  • examples/ingesters/generic-webhook/ — minimal HTTP endpoint
    that accepts POSTs and appends to hive/raw/<source>/<topic>.md.
    Use it to capture context from anything that can send HTTP (Linear,
    GitHub webhooks, Zapier, cron, internal tools).
  • examples/silo-mature/coder/ — synthetic populated silo showing
    what a mature agent memory looks like after weeks of use. Log,
    context, memory, and learnings files with realistic density.
    Reference content only; not installed.
  • examples/README.md — index explaining what's under
    examples/ and when to use each piece.

Documented (existing behavior, previously implicit)

  • hive/knowledge/ is expandable. Ships with three canonical files
    (HUMAN_CONTEXT, SOUL, DOMAINS); curator adds topical files as the
    hive grows (PROJECTS, COMPLIANCE, per-topic, etc.). Previously
    implicit; now called out in HIVE_ARCHITECTURE.md.
  • hive/learnings/distilled/ is expandable. Ships with four
    canonical files (patterns, mistakes, wins, cross-agent-insights);
    curator adds topical files (e.g. auth-patterns.md) as the hive
    accumulates enough distilled content per topic to warrant its own
    file. Previously implicit; now documented.

Changed

  • Default install is now zero-input. curl | sh scaffolds the
    reserved main silo, wires the managed CLAUDE.md block, and exits.
    No wizard, no prompts. Users opt into the interactive setup with
    MEMORY_HIVE_WIZARD=1 curl | sh or memory-hive setup after install.
  • Heuristic role matching in the installer now uses only generic
    patterns (*-coder, *-dev, *-engineer, *-auditor, *-analyst,
    *-strategist, *-specialist, *-pm, etc.). Specific agent names
    from any individual user's roster have been removed.
  • hive/knowledge/HUMAN_CONTEXT.md is now a blank template. Each
    user fills in their own facts locally; nothing personal ships in the
    public repo.

Added

  • memory-hive setup — CLI verb that re-invokes the installer with
    MEMORY_HIVE_WIZARD=1 on demand. Useful after a silent install when
    you decide you want the guided flow.
Assets 2
Loading