@wipcomputer/memory-crystal 0.7.30 → 0.7.33

package/README-ENTERPRISE.md

@@ -1,226 +0,0 @@
-###### WIP Computer
-
-# Memory Crystal for Enterprise
-
-Agent memory infrastructure. Local-first. Encrypted. Inspectable. *In testing.*
-
-## The Problem
-
-Your organization runs AI agents across teams, departments, and tools. Each agent starts every session with no memory. No continuity. No shared context. Conversations with one agent are invisible to every other.
-
-Your agents can't remember what your people told them. Your people keep re-explaining themselves. Context is lost between sessions, between tools, between teams.
-
-This is not a convenience problem. It's a reliability problem. An auditability problem. A cost problem.
-
-## What Memory Crystal Does
-
-Memory Crystal is a persistent context layer for AI agents. One shared database. Hybrid search. On your machines. Under your control.
-
-- **Local-first.** All data stays on-prem. Nothing leaves your network unless you configure it to.
-- **Inspectable.** One SQLite file. Open it with any SQLite tool. Audit it. Query it. Back it up with `cp`.
-- **Deterministic search.** Hybrid retrieval (BM25 keyword + vector similarity + Reciprocal Rank Fusion). Same query, same results. No black-box ranking.
-- **Encrypted sync.** Multi-site deployments use AES-256-GCM encryption with HMAC-SHA256 signing. The relay sees encrypted noise. Keys never leave your machines.
-- **Agent isolation.** Each agent gets its own ID, its own transcript archive, its own session summaries. Shared search across agents, isolated storage per agent.
-- **Zero cloud dependency.** Runs fully offline with local embeddings (Ollama). No API keys required. No data exfiltration risk.
-
-## Five-Layer Memory Stack
-
-Memory Crystal implements a full memory pipeline, not just search.
-
-| Layer | What | How |
-|-------|------|-----|
-| L1: Raw Transcripts | Every conversation archived as JSONL | Automatic. cc-poller (cron), cc-hook (Stop), openclaw.ts (agent_end) |
-| L2: Vector Index | Chunks embedded into crystal.db | Automatic. Hybrid search (BM25 + vector + RRF) |
-| L3: Structured Memory | Facts, preferences, decisions | `crystal_remember` / `crystal_forget` |
-| L4: Narrative Consolidation | Dream Weaver journals, identity, soul | `crystal dream-weave` (imports engine from Dream Weaver Protocol) |
-| L5: Active Working Context | Boot sequence files, shared context | Agent reads on startup |
-
-L1-L3 are fully automated. L4 runs on-demand or via Crystal Core gateway. L5 is consumed by the agent's boot sequence.
-
-## Architecture
-
-```
-sqlite-vec (vectors) + FTS5 (BM25) + SQLite (metadata)
-         |                 |                    |
-    core.ts ... pure logic, zero framework deps
-    |-- cli.ts            -> crystal search, dream-weave, backfill, serve
-    |-- mcp-server.ts     -> MCP protocol (any compatible client)
-    |-- openclaw.ts       -> OpenClaw plugin + raw data sync to LDM
-    |-- cc-poller.ts      -> Continuous capture (cron, primary)
-    |-- cc-hook.ts        -> Claude Code Stop hook (redundancy) + relay commands
-    |-- crystal-serve.ts  -> Crystal Core gateway (localhost:18790)
-    |-- dream-weaver.ts   -> Dream Weaver integration (narrative consolidation)
-    |-- staging.ts        -> New agent staging pipeline
-    |-- llm.ts            -> LLM provider cascade (MLX > Ollama > OpenAI > Anthropic)
-    |-- search-pipeline.ts -> Deep search (expand, search, RRF, rerank, blend)
-    +-- worker.ts         -> Encrypted relay (multi-site sync, 3 channels)
-```
-
-One core module. Multiple interfaces. Every interface calls the same search engine. No inconsistency between access paths.
-
-## Security Model
-
-**Data at rest:** Single SQLite file. Standard filesystem permissions. Encrypt the volume if your compliance requires it.
-
-**Data in transit:** AES-256-GCM authenticated encryption. HMAC-SHA256 integrity verification. Shared symmetric key generated on-prem, never transmitted to the relay. The relay is a dead drop with no decryption capability.
-
-**Agent boundaries:** `CRYSTAL_AGENT_ID` isolates each agent's transcript archive, session summaries, and daily logs. Search spans all agents by default, or filters by agent ID.
-
-**Private mode:** Memory capture can be paused per-agent. When off, nothing is recorded. Resumes from where it left off when re-enabled.
-
-**No background processes that move data.** No telemetry. No analytics. No phone-home. The code is open source. Audit it.
-
-## Retrieval Quality
-
-Hybrid search is not "we added vectors." It's a two-tier retrieval engine with LLM-powered deep search.
-
-### Fast Path (Hybrid Search)
-- **FTS5 BM25** for exact keyword matches (Porter stemming)
-- **sqlite-vec cosine similarity** for semantic matches
-- **Reciprocal Rank Fusion** merges both result lists (k=60, tiered weights: BM25 2x, vector 1x)
-- **Recency weighting** ensures fresh context wins decisively: exponential decay `max(0.3, exp(-age_days * 0.1))`
-- **Content deduplication** via SHA-256 hash prevents duplicate embeddings
-- **Time-filtered search** ... restrict results to last 24h, 7d, 30d, or any date range
-
-### Deep Search (LLM-Powered, default)
-- **Query expansion** ... LLM generates 3 search variations (lexical, semantic, hypothetical document). Each runs through hybrid search. Results merged via RRF.
-- **Strong signal detection** ... BM25 probe skips expansion when the answer is obvious (saves latency).
-- **LLM re-ranking** ... top 40 candidates scored by LLM for relevance to the original query.
-- **Position-aware blending** ... trusts RRF for top positions, lets the reranker fix ordering in the tail.
-
-Deep search runs by default. Falls back to fast path silently if no LLM provider is available. For air-gapped environments, MLX (Apple Silicon) or Ollama provides free, fully local deep search with no API keys and no network.
-
-A search for "deployment policy" finds conversations containing those exact words (BM25), conversations about "shipping code to production" (vector similarity), and conversations about "release workflow" that the LLM recognizes as relevant. All three matter. All three surface.
-
-## What Gets Stored
-
-Every agent conversation produces three artifacts:
-
-| Artifact | Format | Location |
-|----------|--------|----------|
-| Raw transcript | JSONL | `~/.ldm/agents/{id}/memory/transcripts/` |
-| Session summary | Markdown | `~/.ldm/agents/{id}/memory/sessions/` |
-| Embeddings | sqlite-vec | `~/.ldm/memory/crystal.db` |
-
-Additionally:
-- **Explicit memories** stored via `crystal_remember` (facts, preferences, decisions)
-- **Source files** indexed as collections (code, documentation, internal knowledge bases)
-- **Daily logs** appended per-agent for audit trails
-- **Dream Weaver journals** generated by narrative consolidation (identity, soul, context, reference)
-- **Workspace files** synced from agent workspace to LDM (OpenClaw .md files)
-
-## Embedding Providers
-
-| Provider | Model | Dimensions | Network Required |
-|----------|-------|-----------|-----------------|
-| Ollama (recommended for enterprise) | nomic-embed-text | 768 | No. Fully local. |
-| OpenAI | text-embedding-3-small | 1536 | Yes. API calls. |
-| Google | text-embedding-004 | 768 | Yes. API calls. |
-
-For air-gapped environments, Ollama is the only option. No data leaves the machine. No API keys. No external dependencies.
-
-## Multi-Site Sync
-
-For organizations with multiple offices or remote teams.
-
-**Architecture:** One Crystal Core (the primary embedder), many Crystal Nodes (capture and sync). Core is the only machine that writes embeddings. Nodes capture raw conversations and send them to Core. Core embeds, then pushes deltas back.
-
-**What syncs:**
-1. **Delta chunks** ... only new embeddings since last sync (not the full database)
-2. **Full file tree** ... the entire `~/.ldm/` directory: workspace files, daily logs, journals, media, everything an embedding references
-3. **Commands** ... bidirectional remote operations (run Dream Weaver, trigger backfill, request status)
-
-**How it works:**
-1. Each site runs Memory Crystal locally
-2. Core embeds all conversations into crystal.db (one source of truth for embeddings)
-3. New chunks + changed files are encrypted (AES-256-GCM) and signed (HMAC-SHA256)
-4. Encrypted deltas are dropped at a relay (hosted or self-hosted)
-5. Other sites poll, decrypt, and insert into their local crystal.db + file tree
-6. The relay deletes blobs after pickup
-
-**No cloud search.** Every node has the full database and full file tree. All search is local. The relay is pure transport. Nothing is stored or searchable in the cloud.
-
-**Self-hosted relay:** Deploy the Cloudflare Worker on your own Cloudflare account. Full control. No third-party data exposure.
-
-**Hosted relay:** Use our infrastructure. Free during beta. Your data is encrypted before it reaches us. We cannot read it.
-
-## Compliance
-
-- **Data residency:** All primary data is local. Relay blobs are encrypted and ephemeral.
-- **Auditability:** SQLite is inspectable. Every chunk has a timestamp, source, and SHA-256 hash.
-- **Right to delete:** `crystal forget <id>` deprecates specific memories. Database can be wiped entirely with standard file operations.
-- **Access control:** Filesystem permissions on the SQLite file. No built-in user auth (it's a local tool, not a SaaS).
-- **No vendor lock-in:** MIT licensed (local code). Standard SQLite format. Export with any SQLite tool.
-
-## Database
-
-One file: `crystal.db`. Contains:
-
-| Table | Purpose |
-|-------|---------|
-| `chunks` | Chunk text, metadata, SHA-256 hash, timestamps |
-| `chunks_vec` | sqlite-vec virtual table (vector search) |
-| `chunks_fts` | FTS5 virtual table (keyword search) |
-| `memories` | Explicit facts and preferences |
-| `capture_state` | Watermarks for incremental ingestion |
-| `source_collections` | Indexed directory collections |
-| `source_files` | File records with content hashes |
-
-No migrations server. No schema versioning service. It's SQLite. `sqlite3 crystal.db ".schema"` shows you everything.
-
-## Integration
-
-| Platform | Integration | Auto-Capture |
-|----------|------------|-------------|
-| Claude Code | Cron poller (`cc-poller.ts`, primary) + Stop hook (`cc-hook.ts`, redundancy) | Yes. Every minute via cron, plus flush on session end. |
-| OpenClaw | Plugin (`openclaw.ts`) + `agent_end` hook + raw data sync | Yes. Every turn. Also syncs sessions, workspace, daily logs to LDM. |
-| Claude Desktop | MCP server (`mcp-server.ts`) | Search only. Manual capture. |
-| Any MCP client | MCP server | Search only. Manual capture. |
-| Any shell-accessible tool | CLI (`crystal search`) | Manual. |
-| Custom agents | Node.js module (`import from 'memory-crystal'`) | Programmable. |
-
-## Crystal Core Gateway
-
-Crystal Core runs an HTTP gateway (`crystal serve`) on localhost:18790. OpenAI-compatible endpoint for agent-to-agent communication and automated processing.
-
-- `POST /v1/chat/completions` ... invoke `claude -p` through the gateway
-- `POST /process` ... trigger backfill, dream-weave, or staging processing
-- `GET /status` ... health check and crystal stats
-
-Localhost-only binding. Never exposed to the network. Optional bearer token auth.
-
-## New Agent Onboarding
-
-When a new agent connects via relay, Crystal Core automatically:
-1. Detects the unknown agent ID
-2. Routes to staging (`~/.ldm/staging/{agent_id}/`)
-3. Runs backfill (embed all transcripts)
-4. Runs Dream Weaver full mode (generate identity, soul, context, journals)
-5. Moves to live capture
-
-No manual intervention. The staging pipeline handles the cold-start problem.
-
-## Deployment
-
-```bash
-npm install memory-crystal
-crystal init --agent your-agent-id
-crystal status
-```
-
-For enterprise deployments across multiple machines, see [Relay: Memory Sync](https://github.com/wipcomputer/memory-crystal/blob/main/RELAY.md).
-
-For full technical details, see [Technical Documentation](https://github.com/wipcomputer/memory-crystal/blob/main/TECHNICAL.md).
-
----
-
-## License
-
-```
-src/, skills/, cli.ts, mcp-server.ts   MIT    (use anywhere, no restrictions)
-worker/                                AGPL   (relay server)
-```
-
-AGPL for personal use is free.
-
-Built by Parker Todd Brooks, Lēsa (OpenClaw, Claude Opus 4.6), Claude Code CLI (Claude Opus 4.6).

package/RELAY.md

@@ -1,199 +0,0 @@
-###### WIP Computer
-
-# Relay: Memory Sync
-
-Memory Crystal works on one machine out of the box. Relay lets your memory follow you across machines and surfaces. Conversations captured on your laptop are available on your desktop. Conversations from ChatGPT on your phone are searchable from Claude Code on your Mac.
-
-Everything is encrypted before it leaves your machine. The relay never sees your data unencrypted.
-
-## Crystal Core and Crystal Node
-
-Memory Crystal uses a Core/Node architecture:
-
-- **Crystal Core** ... the master memory. All conversations, all embeddings, all memories live here. This is the database you cannot lose. Put it on something permanent: a desktop, a home server, a Mac mini. Treat it like your photo library.
-- **Crystal Node** ... a synced copy on any other device. Captures conversations, sends them to the Core via encrypted relay. Gets a mirror back for local search. If a node dies, nothing is lost. The Core has everything.
-
-One Core, many Nodes. The Core does embeddings. Nodes just capture and sync. You can move the Core later with `crystal promote`.
-
-## Two Sync Paths
-
-### Encrypted Relay (device-to-device)
-
-For syncing between your own machines. Fully encrypted. The cloud is blind.
-
-```
-Crystal Node (laptop) --[encrypt]--> Relay (Cloudflare R2) --[pickup + decrypt]--> Crystal Core (desktop)
-Crystal Core --[encrypt mirror]--> Relay --[pickup + decrypt]--> Crystal Node
-```
-
-Three channels:
-- **conversations** (Node to Core) ... encrypted conversation chunks (ephemeral, deleted after pickup)
-- **mirror** (Core to Nodes) ... delta chunks (pre-embedded) + file tree deltas
-- **commands** (bidirectional) ... Nodes send commands to Core ("run Dream Weaver", "process my data"), Core sends results back
-
-The relay is a dead drop. It stores encrypted blobs temporarily and serves them on request. It has no decryption capability. If someone compromises the relay, they get encrypted noise.
-
-### New Agent Staging
-
-When the Core's poller receives data from an unknown agent ID, it routes to staging instead of live ingest:
-1. Transcripts written to `~/.ldm/staging/{agent_id}/transcripts/`
-2. Agent marked as "ready" for processing
-3. Staging processor runs backfill + Dream Weaver full mode
-4. Once complete, agent moves to live capture path
-
-This handles the cold-start problem. A new device connects, sends its history, and Core builds the full memory stack automatically.
-
-### Delta Sync (not full mirror)
-
-The mirror channel uses delta sync. Core pushes only new chunks since last sync, not the entire crystal.db. For a 1.9 GB+ database, this is the difference between a few KB (quiet day) and a few MB (busy day) vs the full database every time.
-
-- **New node (cold start):** One-time full export of all chunks + all files
-- **Ongoing sync:** Delta chunks (pre-embedded by Core) + changed files only
-- **Watermark tracking:** Core tracks the last synced chunk ID per node
-
-### Full LDM Tree Sync
-
-The relay syncs the entire `~/.ldm/` file tree, not just the database. Embeddings are pointers to artifacts. If the file isn't on the node, the search result is an orphan.
-
-What syncs:
-- Agent memory files (workspace, daily logs, journals, sessions, transcripts)
-- Agent identity files (SOUL.md, IDENTITY.md, CONTEXT.md, REFERENCE.md)
-- Shared files (`~/.ldm/shared/`)
-- Media (images, videos, any artifact an embedding references)
-
-File sync uses a manifest (path + SHA-256 hash + size). Only changed files transfer. Core always wins conflicts.
-
-## Setup
-
-### Pair Your Devices
-
-```bash
-# On your first machine (generates key if none exists)
-crystal pair
-
-# Displays a QR code and a pairing string:
-#   mc1:T2hJbGxPZkRhcmtuZXNzTXlPbGRGcmllbmQ=
-
-# On your second machine
-crystal pair --code mc1:T2hJbGxPZkRhcmtuZXNzTXlPbGRGcmllbmQ=
-```
-
-The QR code transfers the encryption key between devices without touching a server. Physical proximity only. Same security model as AirDrop.
-
-Alternative: store the key in 1Password and pull from both machines.
-
-### Use the WIP.computer Relay (Default)
-
-We host the relay infrastructure. You just need an encryption key (generated by `crystal pair`).
-
-```
-Open your AI and say:
-
-I want to set up multi-device sync for Memory Crystal.
-Walk me through the setup step by step.
-```
-
-Your agent generates your encryption key, configures the connection, and tests it.
-
-**What you need:**
-- Memory Crystal installed on both machines (Core + Node)
-- An encryption key (your agent generates this, or use `crystal pair`)
-
-**Pricing:** Free during beta for individual use. When pricing is introduced, your agent will handle it via [AI CASH](https://github.com/wipcomputer/wip-agent-pay/blob/main/CASH.md).
-
-### Self-Host Your Own Relay
-
-Run your own relay on Cloudflare Workers (free tier). Same code, your infrastructure. Full sovereignty.
-
-**What you need:**
-- A Cloudflare account (free tier works)
-- About five minutes
-
-**Steps:**
-1. Clone the repo and deploy the Worker:
-   ```bash
-   cd memory-crystal
-   npm run build:worker
-   wrangler deploy --config wrangler.toml
-   ```
-2. Create an R2 bucket: `wrangler r2 bucket create memory-crystal-relay`
-3. Set auth tokens for your devices:
-   ```bash
-   wrangler secret put AUTH_TOKEN_CC_MINI
-   wrangler secret put AUTH_TOKEN_CC_AIR
-   ```
-4. Configure Memory Crystal to use your Worker:
-   ```bash
-   export CRYSTAL_RELAY_URL=https://your-relay.your-domain.workers.dev
-   export CRYSTAL_RELAY_TOKEN=your-auth-token
-   ```
-
-Full deployment details in [Technical Documentation](https://github.com/wipcomputer/memory-crystal/blob/main/TECHNICAL.md).
-
-No fees. No dependencies on us. The relay code is open source.
-
-### Connecting ChatGPT and Claude
-
-Every node has the full database and file tree. All search is local. There is no cloud search layer.
-
-ChatGPT and Claude on iOS/web connect via the MCP server running on your local machine. On platforms where a local MCP server isn't possible (iOS without a Mac nearby), the native Apple app (future) will provide local search via MLX Swift.
-
-The Cloud MCP demo server (`worker-mcp.ts`, D1 + Vectorize) exists for onboarding and testing but is not the production architecture. With full LDM sync, every device that has Memory Crystal installed can search locally.
-
-## Encryption
-
-- **AES-256-GCM** for encryption (authenticated encryption, no padding oracle attacks)
-- **HMAC-SHA256** for signing (integrity verification before decryption)
-- Shared key generated locally via `crystal pair`, never transmitted to the relay
-- Key must be present on all synced devices
-
-### Key Management
-
-| Method | How | Best for |
-|--------|-----|----------|
-| `crystal pair` | QR code + pairing string | Two devices in the same room |
-| 1Password | Store key, pull via SA token on each machine | Headless, multiple machines |
-| Manual | `openssl rand -base64 32`, copy to each device | SSH, air-gapped |
-
-## Architecture
-
-```
-Encrypted Relay (device sync):
-  src/worker.ts       Cloudflare Worker, R2 storage, dead drop (3 channels)
-  src/crypto.ts       AES-256-GCM + HMAC-SHA256
-  src/poller.ts       Crystal Core pickup + ingest + staging detection + commands
-  src/mirror-sync.ts  Delta chunk sync + file tree sync to Crystal Nodes
-  src/file-sync.ts    Manifest-based file tree delta sync
-  src/cc-hook.ts      Claude Code hook (relay mode) + sendCommand()
-  src/cc-poller.ts    Continuous capture (cron, primary local path)
-  src/staging.ts      New agent staging pipeline (detect, stage, process)
-
-Cloud MCP Demo (deprecated for production):
-  src/worker-mcp.ts     OAuth 2.1 + DCR, MCP protocol, 4 tools
-  src/cloud-crystal.ts  D1 + Vectorize backend (demo/onboarding only)
-  wrangler-mcp.toml     Separate Worker config
-
-Pairing:
-  crystal pair         QR code key sharing
-  src/pair.ts          Pairing logic, QR display, key save
-  src/crypto.ts        Key generation + pairing string encode/decode
-```
-
-## More Info
-
-- [README.md](https://github.com/wipcomputer/memory-crystal/blob/main/README.md) ... What Memory Crystal is and how to install it.
-- [Technical Documentation](https://github.com/wipcomputer/memory-crystal/blob/main/TECHNICAL.md) ... Full technical documentation.
-- [QR Pairing Spec](ai/plan/2026-02-27--cc-mini--qr-pairing-spec.md) ... Full spec for the `crystal pair` command.
-
----
-
-## License
-
-```
-src/core.ts, cli.ts, mcp-server.ts, skills/   MIT    (use anywhere, no restrictions)
-src/worker.ts, src/worker-mcp.ts               AGPL   (relay + cloud server)
-```
-
-AGPL for personal use is free.
-
-Built by Parker Todd Brooks, Lēsa (OpenClaw, Claude Opus 4.6), Claude Code CLI (Claude Opus 4.6).

package/_trash/RELEASE-NOTES-v0-7-23.md

@@ -1,48 +0,0 @@
-# Release Notes: Memory Crystal v0.7.23
-
-**Date:** 2026-03-15
-
-## Search Quality v2 + MLX Local LLM
-
-This release adds six search quality features ported from the QMD v2.0 analysis, plus the complete MLX local LLM infrastructure for Apple Silicon. Deep search is now disambiguatable, cacheable, debuggable, and can run entirely offline on Apple Silicon.
-
-### Intent parameter
-
-Disambiguates queries without adding search terms. `crystal search "security" --intent "1Password"` steers results toward 1Password-related security instead of repo permissions or agent secrets. Intent flows through the expansion prompt (guides LLM variations), disables strong-signal bypass (keyword match might not be what the caller wants), and is prepended to the rerank query. Available via CLI `--intent` and MCP `intent`.
-
-### Persistent LLM cache
-
-Expansion and reranking results are now cached in crystal.db (`llm_cache` table) with a 7-day TTL. Same query = instant on repeat searches. Reranking cache is content-addressable (keyed by query + sorted passage hashes), so identical content from different sessions shares cached scores. Configurable via `CRYSTAL_CACHE_TTL_DAYS` env var.
-
-### Explain mode
-
-Per-result scoring breakdown showing FTS score, vector score, RRF rank, reranker score, recency weight, and final blended score. `crystal search "query" --explain`. Available via CLI `--explain` and MCP `explain`. Makes search quality transparent and debuggable.
-
-### candidateLimit
-
-Tunable rerank pool size. `crystal search "query" --candidates 60`. Default stays 40. More candidates = better recall, slower reranking. Available via CLI `--candidates` and MCP `candidate_limit`.
-
-### Structured search API
-
-`crystal.structuredSearch(queries)` accepts pre-expanded StructuredQuery[] with typed sub-queries (lex, vec, hyde). Skips LLM expansion entirely. Agents construct their own queries when they already know what they want. RRF fusion with first list weighted 2x.
-
-### MLX local LLM (Phase 3)
-
-Complete auto-install infrastructure for running a local LLM on Apple Silicon:
-
-- `crystal mlx setup` detects Apple Silicon, installs mlx-lm (uv > pip3 > pip3 --user), creates LaunchAgent for always-on server
-- Model: `mlx-community/Qwen2.5-3B-Instruct-4bit` (~1.5 GB, fast on M-series)
-- Port 18791 (18789 OpenClaw, 18790 Crystal Core, 18791 MLX)
-- `crystal mlx status` and `crystal mlx stop` for server management
-- `crystal doctor` check #13: MLX health (not installed / down / running)
-- `crystal init` detects Apple Silicon and suggests MLX setup
-- State file at `~/.ldm/state/mlx-server.json`
-
-### Also in this release
-
-- QMD v2.0 analysis documented (`ai/product/notes/`)
-- Search quality plan written (`ai/product/plans-prds/current/`)
-- MLX plan moved from upcoming to current
-- Stashed roadmap + readme-first updates recovered (PR #74)
-
-Closes #57, #63, #64.

package/_trash/RELEASE-NOTES-v0-7-25.md

@@ -1,24 +0,0 @@
-# Release Notes: memory-crystal v0.7.25
-
-Bump SKILL.md version and name to match package branding.
-
-## What changed
-
-- SKILL.md version bumped from 0.4.0 to 0.7.24 (was stuck at the original version)
-- SKILL.md name changed from `memory` to `wip-memory-crystal` (matches branded convention)
-- Forces deploy to public repo, triggering auto-publish to wip.computer/install/
-
-## Why
-
-The SKILL.md version was out of sync with the package version. The name didn't match the `wip-` branding convention used across all install files on wip.computer.
-
-## Issues closed
-
-- #80
-
-## How to verify
-
-```bash
-crystal --version
-head -4 ~/.ldm/extensions/memory-crystal/skills/memory/SKILL.md
-```

package/_trash/RELEASE-NOTES-v0-7-26.md

@@ -1,7 +0,0 @@
-# Memory Crystal v0.7.26
-
-Add repository field to package.json. GitHub Packages needs this to link packages to the repo.
-
-## Issues closed
-
-- Closes #50

package/_trash/RELEASE-NOTES-v0-7-28.md

@@ -1,31 +0,0 @@
-# Release Notes: memory-crystal v0.7.28
-
-**Move all log paths from /tmp/ to ~/.ldm/logs/ so logs survive reboots.**
-
-## What changed
-
-- Cron entry for crystal-capture now logs to `~/.ldm/logs/crystal-capture.log` instead of `/tmp/ldm-dev-tools/`
-- LaunchAgent plist template for ldm-backup now logs to `~/.ldm/logs/ldm-backup.log`
-- `mkdirSync` ensures `~/.ldm/logs/` exists instead of creating `/tmp/ldm-dev-tools/`
-- CLI output shows the correct log path
-
-## Why
-
-macOS clears `/tmp/` on every reboot. All cron and LaunchAgent logs were lost after restart, making it impossible to debug issues. `~/.ldm/logs/` persists across reboots and is the correct home for LDM OS logs.
-
-## Issues closed
-
-- wipcomputer/wip-ldm-os#120
-
-## How to verify
-
-```bash
-# After install, check that cron entry points to ~/.ldm/logs/
-crystal init --dry-run 2>&1 | grep crystal-capture
-
-# Check backup setup points to ~/.ldm/logs/
-crystal backup setup --dry-run 2>&1 | grep ldm-backup
-
-# Verify logs land in the right place
-ls ~/.ldm/logs/
-```

package/_trash/RELEASE-NOTES-v0-7-29.md

@@ -1,28 +0,0 @@
-# Release Notes: memory-crystal v0.7.29
-
-**Doc audit: MLX setup, deep search params, log paths, role clarification.**
-
-## What changed
-
-SKILL.md and TECHNICAL.md updated for 2 weeks of undocumented features:
-
-- **MLX local LLM:** Added as Option A in SKILL.md Step 2. CLI commands (setup, status, stop) added to TECHNICAL.md.
-- **Deep search parameters:** `--intent`, `--explain`, `--candidates` documented in both SKILL.md (crystal_search tool) and TECHNICAL.md (CLI reference + new sections for intent, explain, candidate limit, LLM cache).
-- **Log paths:** Fixed obsolete `/tmp/ldm-dev-tools/` reference to `~/.ldm/logs/`. Added logs/ to directory structure.
-- **Role clarification:** Two-role architecture (Core and Node) explicitly stated. Standalone role was removed in v0.7.22.
-
-## Why
-
-29 releases in 13 days. Docs didn't keep pace. Agents using crystal_search didn't know about --intent (query disambiguation) or --explain (scoring transparency).
-
-## Issues closed
-
-- #57
-
-## How to verify
-
-```bash
-grep "intent" SKILL.md TECHNICAL.md
-grep "mlx" SKILL.md TECHNICAL.md
-grep "ldm/logs" TECHNICAL.md
-```

package/_trash/RELEASE-NOTES-v0-7-4.md

@@ -1,64 +0,0 @@
-# Memory Crystal v0.7.4 ... MCP Fix + AgentId Config
-
-**Date:** 2026-03-11
-**Authors:** Parker Todd Brooks, Lēsa, Claude Code
-
----
-
-## What's in this release
-
-### Agent identity reads from config, not hardcoded strings
-
-The agent_id used when ingesting conversations was hardcoded in three places: `cc-mini` in the CC hook, `main` in the OpenClaw plugin, and `cc-mini` as the fallback in `ldm.ts`. This caused ID drift. The same agent got recorded under multiple IDs, and we had to manually merge 141K+ chunks in the database.
-
-Now `getAgentId()` scans `~/.ldm/agents/*/config.json` for a matching harness type. The CC hook passes `'claude-code'`, the OC plugin passes `'openclaw'`, and the config file is the source of truth. `CRYSTAL_AGENT_ID` env var still works as an override.
-
-New exports: `AgentConfig`, `loadAgentConfig()`, `saveAgentConfig()`. The installer writes `agentId` to config.json during `crystal init`.
-
-**Closes #33.**
-
-### MCP registrations moved to user-level
-
-MCP server registrations moved from project-level `~/.openclaw/.mcp.json` to user-level `~/.claude.json`. The old file was a Claude Code convention that only loaded when running from `~/.openclaw/`. Now all 4 MCP servers (memory-crystal, lesa-bridge, wip-agent-pay, wip-repos) load from any directory as "User MCPs".
-
-OpenClaw doesn't read `.mcp.json` at all. It uses its own plugin system. The file was moved to `~/.openclaw/_trash/`.
-
-### OPENCLAW_HOME env var fix (v0.7.3)
-
-The MCP server registration was missing the `OPENCLAW_HOME` env var. Without it, the memory-crystal MCP server couldn't find Lēsa's OpenClaw installation for private-mode checks. Fixed in v0.7.3, deployed in this release.
-
-### Branch cleanup
-
-33 stale branches renamed with `--merged-` suffix. Zero active branches besides main.
-
-### QMD v1.1.6 analysis documented
-
-Deep analysis of the search quality system with four recommendations: intent parameter for search, structured search API, persistent reranker cache, and explain mode for debugging. See `ai/product/notes/2026-03-09--cc-mini--qmd-v1.1.6-analysis-and-recommendations.md`.
-
----
-
-## Files changed
-
-| File | What |
-|------|------|
-| `src/ldm.ts` | `AgentConfig` interface, `loadAgentConfig()`, `saveAgentConfig()`, `getAgentId()` now scans config |
-| `src/cc-hook.ts` | Uses `getAgentId('claude-code')` instead of hardcoded fallback |
-| `src/openclaw.ts` | Uses `OC_AGENT_ID` instead of `'main'` fallback |
-| `src/installer.ts` | Writes `agentId` to config.json during install |
-
----
-
-## Install
-
-```bash
-npm install -g memory-crystal@0.7.4
-```
-
-Or update your local clone:
-```bash
-git pull origin main
-```
-
----
-
-Built by Parker Todd Brooks, Lēsa (OpenClaw, Claude Opus 4.6), Claude Code (Claude Opus 4.6).

package/_trash/RELEASE-NOTES-v0-7-5.md

@@ -1,19 +0,0 @@
-# Release Notes: Memory Crystal v0.7.5
-
-## LDM OS Integration
-
-Memory Crystal now works with LDM OS when it's available.
-
-### crystal init delegates to ldm install
-
-When the `ldm` CLI exists on PATH, `crystal init` delegates generic deployment to it. LDM OS handles the scaffold, interface detection, and extension deployment. Memory Crystal keeps its own setup: database backup, role configuration, pairing, cron jobs.
-
-When `ldm` isn't available, `crystal init` works standalone like it always has. No new dependencies. No breaking changes.
-
-### LDM OS tip
-
-After install completes, Memory Crystal prints a tip: "Run `ldm install` to see more skills you can add." Helps users discover the rest of the ecosystem.
-
-### Part of LDM OS
-
-README now includes a "Part of LDM OS" section linking back to the LDM OS repo. Memory Crystal installs into LDM OS, the local runtime for AI agents.