CHANGELOG.md

MATHIR Changelog

[8.9.6] — 2026-07-31 — CODEX MCP STDIO BANNER FIX + INTEGRATION GUIDE

Fixed

• mathir_mcp_server.py:1490 — mcp.run() → mcp.run(show_banner=False). FastMCP 3.4.4 prints a large colored ASCII banner to stdout at startup; for MCP-over-stdio this corrupts JSON-RPC framing and any MCP host reading JSON-RPC straight off stdout silently fails to register any tool (Codex, in particular). stderr is unaffected, so daemon logging and 3-layer cache logs still appear. First symptom reported live: "MCP server mathir is configured but mcp__mathir__* tools never appear in Codex's tool list." Discovered and patched by Codex's own session, then propagated to source repo per guardrail-sync-deployed-daemon.

Notes

• This is the third layer of three separate MCP-server-for-Codex failures encountered on 2026-07-31 (the other two were ~ not expanded in MCP env vars, and Codex's child process not inheriting python on PATH).
• Documentation: docs/troubleshooting/codex-mcp-tools-invisible-three-layers.md.

[8.9.5] — 2026-07-21 — AUTONOMOUS MAINTENANCE + HEADLESS GOD-MODE WORKERS

Added

• Autonomous "sleep" maintenance thread (_maintenance_loop in mathir_server.py) — a background daemon thread that periodically runs run_maintenance() (decay/promote/dedupe/link-build) on every DB currently cached in _vec_cache. Previously run_maintenance() existed but nothing ever invoked it automatically — lifecycle transitions only happened if a human or agent explicitly called it. Config-driven via a new "maintenance" block in mathir.json (enabled, interval_hours, do_decay, do_promote, do_dedupe, do_links), with MATHIR_MAINTENANCE_ENABLED/MATHIR_MAINTENANCE_INTERVAL_HOURS env overrides.
• Headless, on-demand god-mode workers — bin/god/god_mode_start.py / god_mode_stop.py launch/kill a headless coding-agent CLI (opencode, mimo, claude, openclaude, codex, gemini, aider, cursor-agent, copilot) as a detached background process, human-triggered only, never autostarted. bin/god/god_worker_daemon.py is the actual execution loop: polls /api/god/poll, claims a task, spawns the target CLI with its documented headless flags, streams output live, retries on a silent no-op (process exits 0 but never called memory_save) or a hang past its timeout, then acks the result. Closes the gap where god_bridge.py only notified and needed a human to manually drive execution.
• bin/god/god_mode_report.py — deterministic, LLM-independent text report that reads a project's SQLite DB directly (bypassing /api/memories's project-resolution quirks) and groups results by task, correctly handling multi-target fan-out. Following a real incident (2026-07-21: 3 workers' answers never reached the human because the orchestrator's own relay failed), mathir_god_orchestre()'s built-in prompt guidance now mandates running this after every dispatch round instead of relying on the orchestrator LLM's memory.
• /api/god/ack daemon route — flips a claimed task's label in place (pending→running→completed/failed) instead of creating a duplicate memory per state change, fixing god_poll re-serving the same stale pending task forever.
• Atomic task claiming — /api/god/poll now wraps its SELECT+UPDATE in BEGIN IMMEDIATE/commit, preventing two parallel pollers sharing an agent name from double-claiming the same task.
• memory_save file_path param — persisted into metadata.file_path; new /api/memory/by_path route does a real json_extract(metadata, '$.file_path') LIKE ? SQL filter (replacing the old semantic-recall-then-post-filter approach, which silently missed structurally-tagged-but-not-semantically-similar memories).
• memory_dashboard now calls a genuinely distinct /api/memory/dashboard route (recent activity, guardrail roster, save trend) instead of secretly proxying /api/memory/stats verbatim.
• Self-healing + worker-silence guidance for orchestrators — new prompt sections requiring mathir_god_orchestre() to fix bugs it hits in MATHIR's own code directly (syncing to the deployed daemon path and restarting it, per the existing sync guardrail), with third-party tool bugs only proposed as GitHub issues pending explicit human approval.

Fixed

• created_at_unix decay-eligibility bug — was hardcoded to literal SQL NULL, which via COALESCE(last_recalled_at, created_at_unix, 0) permanently excluded every never-recalled memory from decay eligibility. Fixed to CAST(strftime('%s', created_at) AS REAL).
• Guardrail saves no longer misfire the anomaly detector — memory_save's anomaly check is now skipped for block_type == 'guardrail' saves, since new guardrails almost always describe novel problems (exactly what the Mahalanobis detector is tuned to flag), which previously caused real guardrail saves to get silently reclassified to tier=immunological and vanish from the always-injected GUARDRAILS block.
• memory_recall_quality lexical grounding check — added a requirement that at least one ≥4-char query token literally appear in the top result's text before trusting a "high" (≥0.7 cosine) quality verdict, after a verified case where a nonsense query scored 0.839 purely from embedding-space coincidence.

Changed

• build_links_all threshold raised 0.7 → 0.88 — 0.7 against the real multilingual-e5-small embedding model produced an almost-complete graph (442,890 links from 666 memories), useless as a "related memories" signal.
• god_bridge.py env vars renamed MYCERISE_STATE_DIR/MYCERISE_LOG_FILE → MATHIR_STATE_DIR/MATHIR_LOG_FILE — leftover naming from before MATHIR was extracted into its own standalone project; defaults still resolve under $XDG_CONFIG_HOME, no functional behavior change beyond the var name itself.
• memory_consolidate(dry_run=True) output bounded — new max_results param (default 50) caps preview events to a compact {memory_id_a, memory_id_b, snippet_a, snippet_b, similarity} shape (100-char snippet cap), replacing the old verbose per-pair shape that hit 228K+ chars on ~700 memories.
• /api/memory/export is now file-based — writes the full export to <DATA_DIR>/exports/export_{project_slug}_{timestamp}.json on disk and returns just the file path + count, instead of inlining potentially huge JSON (previously hit 137,102 chars at 755 memories).

---

[8.9.4] — 2026-07-16 — SELF-HEALING DAEMON + UNIVERSAL INJECTION PROXY

Fixed

• Windows healthcheck watchdog required admin, so it silently never installed. setup-autostart.ps1's 5-minute self-heal task needed -InstallHealthcheck + -RunLevel Highest, which most non-admin installs don't have — a daemon that died mid-session (crash, sleep/resume) stayed dead until next logon. The healthcheck doesn't actually need elevation; it's installed by default now (-SkipHealthcheck to opt out), and install_smart.py's Windows autostart path registers it automatically.
• claude_code_hook.py (UserPromptSubmit auto-injection hook) existed but was dead code — never wired into any settings.json, and even when invoked manually it read hook_input["message"] while Claude Code's real payload field is "prompt", so it silently injected nothing. Fixed the field name and wired it into ~/.claude/settings.json; install_smart.py now does this automatically for Claude Code installs.
• Daemon-side prompt-injection sanitizer (_sanitize_for_prompt) was a no-op. s.replace(tok, tok.strip()) does nothing when tok has no surrounding whitespace — true for every token in the list (</mathir-, {{MATHIR_CONTEXT}}, <|, ### ). Confirmed live: a query/memory containing </mathir-auto-injection> passed through unstripped into /api/context's response, which the hook wraps in exactly that tag — a stored-memory or query-based prompt-injection breakout, live in production since the hook above was wired up. Fixed, then unified with mathir_proxy.py's (correct) independent copy into one shared mathir_sanitize.py so the two can't drift apart again.
• mathir_proxy.py's OpenAI-route default target had a double-/v1 bug (https://api.openai.com/v1 + route path /v1/chat/completions → .../v1/v1/chat/completions) — would have 404'd on every real OpenAI-format request the proxy ever forwarded. Never caught because the proxy had never been run end-to-end before this session. Default target is now https://api.openai.com.

Added

• mathir_proxy.py now speaks Anthropic's native /v1/messages (previously OpenAI-compatible /v1/chat/completions only), augmenting the top-level system field — this is the endpoint Claude Code itself calls, so without this route the proxy could not help the tool most people actually use.
• Multi-upstream routing: an optional X-Mathir-Upstream request header lets one proxy process serve requests bound for different providers (Anthropic, OpenAI, OpenRouter, a local model, ...) instead of needing one process per upstream. Validated against an allowlist (~30 known providers + *.openai.azure.com / *.bedrock-runtime.amazonaws.com suffix matching + loopback for any local model server) — falls back to the process's configured default on any unlisted host, never fails open.
• mathir-proxy.service (systemd) and com.mathir.proxy.plist (launchd), mirroring the existing daemon units, so the proxy gets the same Restart=on-failure / KeepAlive self-healing on Linux/macOS. Windows healthcheck now checks both port 7338 (daemon) and 7339 (proxy).

Removed

• mathir_mcp/brain/ — a stale, unimported fork of mathir_watchdog.py, mathir_prime.py, mathir_inject_proxy.py, and mathir_brain.py, superseded by the mathir_lib/ versions (confirmed by diff + mtime) and already flagged as legacy dead weight in tests/test_module_tree.py. Removed the now-stale brain -> bin entry from mathir_sync.py's sync manifest.

[8.9.3] — 2026-07-15 — CROSS-PLATFORM AUTO-START FIX (hardcoded Python paths)

Fixed

• bin/auto_start.bat (Windows) — was hardcoded to %USERPROFILE%\AppData\Local\Programs\Python\Python311\python.exe, which silently failed the daemon launch on any machine using a different Python (e.g. Miniconda — the case that surfaced this bug). Now resolves dynamically: where python on PATH → py launcher → common install locations (Miniconda, Anaconda, WindowsApps, Programs\PythonXXX).
• bin/mathir-daemon.service (Linux/systemd) — ExecStart hardcoded /usr/bin/python3, which doesn't exist on conda-only or minimal setups. Now resolves python3/python via PATH at start time (ExecStart=/bin/sh -c 'exec "$(command -v python3 || command -v python)" ...'), with a fallback PATH covering ~/.local/bin, ~/miniconda3/bin, ~/anaconda3/bin.
• bin/com.mathir.daemon.plist (macOS/launchd) — default interpreter changed from hardcoded /usr/bin/python3 (removed on newer macOS, never used by Homebrew/conda) to /usr/bin/env python3 (PATH resolution at launch), with a documented PATH fallback covering /opt/homebrew/bin.
• INSTALL_FOR_DEV/install_smart.py (_setup_autostart_macos) — the actual code path that renders the deployed plist. Resolution order changed to: venv python (unchanged) → shutil.which("python3") on the installer's own PATH (new — covers Homebrew/conda/pyenv) → /usr/bin/env python3 fallback (new). Previously defaulted straight to the hardcoded /usr/bin/python3 if no venv was found.
• auto_start.sh (Linux/macOS) and bin/mathir_daemon.py (the HTTP-shim launcher) were already correct — no change needed there.

Synced to ~/.config/MATHIR/mathir_mcp/ (verified byte-identical post-sync).

---

[8.9.2] — 2026-07-05 — GOD-MODE CLIENT BRIDGE

Added

• bin/god/god_bridge.py — cross-platform polling daemon for god-mode (3 modes: worker / orchestrator / observer), stdlib-only, no new deps. Beeps + logs when new god:* events arrive.
• bin/god/god_poll.ps1 — Windows PowerShell one-shot poller.
• bin/god/god_poll.sh — POSIX bash one-shot poller.
• bin/god/PROTOCOL.md — full label taxonomy + message flow.
• bin/god/README.md — usage + env vars (MATHIR_DAEMON_URL, MYCERISE_STATE_DIR, MYCERISE_LOG_FILE).
• docs/GOD_MODE.md — new "Client-side tooling (bin/god/)" section explaining modes, quick start, cross-platform.
• Top-level README.md — added client bridge link + bumped to v8.9.1.
• mathir_mcp/README.md — new "God-mode orchestration" section.

Changed

• Daemon /health reports version: 8.9.1 (was 8.5.1) after pip editable install refresh.
• __version__ + pyproject.toml: 8.5.1 → 8.9.1.
• docs/GOD_MODE.md extended with client-side tooling section (was server-only).
• Default tool count docs: 26 tools → 27 tools (mathir_god_agent + mathir_god_orchestre were missing from the count).

Synced across all 3 MATHIR install locations

• D:\SECRET_PROJECT\MATHIR\mathir_mcp\ (github source / truth)
• ~/.config/MATHIR/mathir_mcp/ (egg install used by Claude / MiMo / OpenCode — bin/god/ populated, .mathir/ DB preserved, 780 memories intact)
• ~/.config/mimocode/tools/mathir_mcp/ (v8.4.0 dev copy — 12 drifted files re-synced, bin/god/ populated)

Hygiene

• 4 stale god:task:*:pending entries purged (had completed siblings).
• ~/.config/MATHIR/mathir_mcp-8.5.1.dist-info (stale pip editable-install metadata) refreshed to 8.9.1 via pip install --force-reinstall --no-deps -e ....
• Mycerise scripts/god-mode/ + scripts/mathir/ deleted (were polluting namespace, never belonged in the consumer project).

---

[8.9.1] — 2026-07-05 — DOC CORRECTION (27 tools canonical)

Fixed

• Canonical tool count corrected: 26 → 27 — verified against live MCP tool list.
  • The audit_immunological tool was missing from the v8.9.0 count.
  • All docs, templates, READMEs, and architecture diagrams now consistently say 27 MCP tools.
• Canonical breakdown: 2 auto-injection + 10 basic + 7 lifecycle + 3 advanced + 1 guardrail + 1 immunological + 1 health + 2 god mode = 27.
• block_type declarations everywhere now list all 6 tiers: working_memory | episodic | semantic | procedural | guardrail | immunological.
• LIVING MEMORY (5 TIERS) headers → (6 TIERS) in all 150+ agent/skill files.
• Guardrail row added to all memory-tier tables (was missing in skill/agent injections).
• Version titles updated to v8.9.0 (was stale v8.5.0/v8.6.0/v8.7.0 in some deployments).
• Templates opencode_templates/README.md: "19 MCP tools" → "27 MCP tools", "5-tier model" → "6-tier model".

Synced deployments

• ~/.config/MATHIR/ (mathir_mcp full copy)
• ~/.config/opencode/ (32 agents + 82 skills + 82 skills-global re-injected)
• ~/.config/mimocode/ (32 agents re-injected)
• ~/.claude/CLAUDE.md

---

[8.9.0] — 2026-07-05 — GUARDRAIL TIER

Added

• guardrail tier — 6th memory tier, push-based. Always auto-injected into every /api/context, memory_session_start, memory_context response.
• memory_list_guardrails MCP tool — list all active guardrails
• /api/memory/guardrails daemon route — GET/POST guardrail listing
• list_guardrails() / count_guardrails() methods in VecMemory
• Guardrails are immune to decay and promotion (terminal tier)
• Minimum priority enforced at 8, max 50 per project

Changed

• Total MCP tools: 27 (corrected in 8.9.1; was miscounted as 26 at release)
• Total tiers: 6 (was 5)

---

[8.8.0] — 2026-07-04 — GOD MODE — MULTI-AGENT ORCHESTRATION

Added

• mathir_god_agent MCP tool — worker self-identification + task polling. Call with no args → agent self-assesses honestly (name, capabilities, strengths, weaknesses). Call with profile → register + poll.
• mathir_god_orchestre MCP tool — orchestrator discovers workers with full profiles, decomposes directives, assigns tasks by worker strength.
• mathir_god.py core module — GodProtocol (label encoding), TaskGraph (DAG + cycle detection), WorkerRegistry (capability lookup), WorktreeManager (git worktree lifecycle)
• /api/god/poll daemon route — optimized SQL polling for pending tasks
• /api/god/agents daemon route — list registered workers with introductions
• Built-in helpers — name="help" and directive="help" return full usage guides
• LIKE injection prevention — % and _ escaped in god/poll queries
• 40 new tests in tests/test_god.py
• docs/GOD_MODE.md — full architecture and usage guide

Changed

• Total MCP tools: 25 (was 23)

---

[8.7.0] — 2026-07-03 — 3-LAYER AUTO-CACHE

Added

• 3-layer auto-cache system (mathir_cache.py) for significant performance boost:
  • L1 Embedding Cache — LRU (1024 entries) on _encode_query()/_encode_passage(). Same text → instant lookup (~60ms → <1ms). Deterministic, never expires.
  • L2 Recall Cache — TTL-based (256 entries, 60s TTL) on /api/memory/recall results. Deduplicates identical queries across agents. Invalidated on any write (save/delete/promote/consolidate).
  • L3 Session Cache — Pre-warmed top-20 memories per project (5 min TTL) on /api/context. Session start and context calls return instantly on repeat.
• /api/cache/stats endpoint — hit/miss counters and hit ratio for all 3 layers
• cache field in recall responses — "hit" or "miss" for observability
• Write-through invalidation — invalidate_on_write() called on save, delete, promote, auto_promote, consolidate (non-dry-run)
• 24 new cache tests (LRU eviction, TTL expiry, invalidation, stats, integration)

Changed

• _encode_query() / _encode_passage() now check L1 cache before calling embedder.encode()
• /api/memory/recall checks L2 cache before running vector search
• /api/context checks L3 cache before running search
• Total tests: 122 (was 98)

[8.6.1] — 2026-07-03 — PORTABLE PATHS + CROSS-PLATFORM INSTALL FIX

Fixed

• Eliminated ALL hardcoded machine-specific paths from the entire codebase
  • Removed C:\Users\<username>\... and D:\SECRET_PROJECT\... references (39 occurrences across 8 files)
  • Removed ALL ~/.config/opencode/bin/ legacy paths (~50 occurrences across 18 files)
• Migrated install location from legacy ~/.config/opencode/bin/ to canonical ~/.config/MATHIR/mathir_mcp/
• Daemon launch bug — auto_start.bat was launching mathir_daemon.py (raw TCP socket) instead of mathir_server.py (HTTP/Flask). Clients make HTTP requests, so the TCP daemon accepted connections but never responded. Fixed in all auto-start scripts.
• VBS launcher (auto_start_vbs.vbs) — replaced hardcoded path with WScript.Shell.ExpandEnvironmentStrings("%USERPROFILE%") for portable resolution
• Python scripts (migrate_mathir_schema.py, test_batch_recall.py) — replaced hardcoded paths with os.path.expanduser("~") based resolution
• install_smart.py macOS plist generation — plist paths now use ~/.config/MATHIR/ instead of ~/.config/opencode/
• Added .claude/ and .mimocode/ to .gitignore (local state, must not be committed)
• DB routing — _resolve_db() was creating DBs in global ~/.config/MATHIR/data/projects/ even when a project cwd was provided. Now creates .mathir/mathir.db inside the project directory for new projects.
• DB routing backward-compat — existing 417 databases in global config are still found: routing checks local .mathir/ first, then global, only creates local for genuinely new projects.
• Case-sensitivity — standardized ~/.config/MATHIR (uppercase) everywhere. The lowercase ~/.config/mathir variant would create a separate directory on Linux/macOS.
• Removed legacy ~/.config/opencode fallback from mathir_paths.py

Changed

• All cross-platform scripts (.bat, .ps1, .sh, .service, .plist) now resolve paths from %USERPROFILE% / $HOME / ~
• All install guides (Windows, Linux, macOS) updated to ~/.config/MATHIR/mathir_mcp/ paths
• All agent/command/docs templates (opencode + mimocode) updated to new canonical paths
• GLOBAL_INSTRUCTIONS.md deployed path updated
• DB routing priority: local .mathir/ → global ~/.config/MATHIR/data/ → create local

Files touched (24 files)

• bin/auto_start.bat, bin/auto_start.sh, bin/auto_start_vbs.vbs
• bin/auto_start_helpers.ps1, bin/auto_start_healthcheck.ps1
• bin/start_daemon_background.ps1, bin/setup-autostart.ps1
• bin/mathir-daemon.service, bin/com.mathir.daemon.plist
• dev/migrate_mathir_schema.py, dev/test_batch_recall.py
• INSTALL_FOR_AGENT/INSTALL_WINDOWS.md, INSTALL_FOR_AGENT/INSTALL_LINUX.md, INSTALL_FOR_AGENT/INSTALL_MACOS.md
• INSTALL_FOR_DEV/README.md, INSTALL_FOR_DEV/install_smart.py
• GLOBAL_INSTRUCTIONS.md
• All opencode_templates/ and mimocode_templates/ inject files

[8.6.0] — 2026-07-03 — INT8 QUANTIZATION + MULTI-AGENT BENCHMARK

Added

• INT8 scalar quantization for embedding storage (algo #22) — 4x compression, 0% recall loss
  • _quantize_int8() / _serialize_embedding() — float32 → int8 at store time
  • vec_int8(X'...') SQL function for sqlite-vec INT8 table operations
  • Automatic FLOAT→INT8 migration on first DB access (transparent, zero-downtime)
  • Migrated 410 production DBs: 1.9 GB → 825 MB (2.3x average reduction)
  • Proven 10/10 recall@10 overlap vs float32 (zero ranking degradation)
• Cross-encoder reranking (algo #21) — CrossEncoderReranker class
  • Model: cross-encoder/ms-marco-MiniLM-L-6-v2 (22M params, lazy-loaded)
  • Integrated into HybridSearch.hybrid_search(rerank=True) — fetches k*3, reranks to k
  • +20pp hit@10 on natural-language queries (50% → 70%)
  • rerank parameter added to MCP memory_hybrid_search tool and HTTP API
• Multi-agent shared memory benchmark (benchmarks/08_industry_validation/multi_agent_bench.py)
  • Tests whether MATHIR can make "dumb" (free-tier) models intelligent via shared memory
  • 3 phases: baseline (no memory) → MATHIR-assisted → multi-agent collaboration
  • Result: 0% → 53% average accuracy with MATHIR (+70pp for best agent)
  • Supports OpenCode Zen free models + MiniMax native API + Groq
• LoCoMo benchmark on Groq (Llama 3.3 70B) and OpenCode Zen (mimo/deepseek/nemotron)
  • Groq partial: 51.2% (21/41 judged, TPM-limited)
  • Zen (mimo+deepseek free): 38.8% (67/152 judged)
• e5-small vs e5-large-v2 embedding comparison on fluid mechanics corpus
  • e5-small + rerank beats e5-large (52.9% vs 51.0%) at 47x less encoding cost
• 4 new cross-encoder reranking tests (98 total tests passing)

Fixed

• find_duplicates() — adapted to INT8: queries use vec_int8(X'...') for MATCH
• mathir_server.py hybrid search — adapted vector query to INT8 format
• Groq LLM client — added 413 retry (TPM rate limit masquerades as "Request too large")
• Groq LLM client — added User-Agent header (fixes Cloudflare 403)
• LoCoMo context truncation — MATHIR_BENCHMARK_CONTEXT_MAX_CHARS env var

Changed

• vec_memories table schema: FLOAT[dim] → INT8[dim] (new DBs)
• _serialize_embedding() returns hex string (was bytes blob)
• _deserialize_embedding() decodes int8 → float32 (was float32 → float32)
• Test test_search_include_embeddings_returns_raw_vectors: asserts cosine > 0.99 (was atol=1e-5)

[8.5.1] — 2026-07-01 — BUG FIXES + AUDIT LOG

Added

• memory_audit table + _log_audit() — real audit trail for save/delete/promote/decay/link
• stability column on memories table (Ebbinghaus decay persistence)
• Anomaly threshold recalibrated from 2.0 → 25.0 using production data
• reset_anomaly_state() function for full anomaly detector reset

Fixed

• Embedder "meta tensor" crash: root-caused to PyTorch 2.6 weights_only=True default
• Anomaly state reset: clearing DB row alone doesn't reset a running daemon's cached detector

[8.5.0] — 2026-06-25 — FASTMCP REWRITE + AUTO-INJECTION + MULTI-SESSION

Changed

• MCP server rewritten using FastMCP 3.4.2 (replaces hand-rolled JSON-RPC stdio loop)
• MCP server v3 = thin HTTP proxy to daemon (port 7338) — NO local embedder loading
• 20 MCP tools (2 auto-injection + 10 basic + 7 lifecycle + 1 health check)
• Multi-session safe: multiple OpenCode sessions share ONE daemon embedder (no CUDA conflicts)
• Unified Flask+Waitress server (mathir_server.py) replaces TCP daemon + http.server
• Auto-injection plugin (mathir-auto-inject.ts) injects memories into system prompt
• /api/context endpoint for plugin auto-injection
• memory_session_start + memory_context tools for session context
• Registry-based DB resolution (checks registry → projects dir → CWD → legacy)
• 127.0.0.1 instead of localhost (Windows IPv6 resolution delay)
• Dependencies: added fastmcp>=3.4.0, removed aiohttp, pyzmq (no longer needed)

Fixed

• PyTorch 2.6 meta tensor crash ("Cannot copy out of meta tensor; no data!")
• Multi-session CUDA crash — root cause was 2+ MCP servers each loading embedder on GPU
• Missing import threading in MCP server (crash on startup)
• Hardcoded path in get_project_db_path
• NULL embeddings (17 memories saved via direct sqlite had NULL vectors)

Security

• Input length caps: content 100KB, query 5KB, label 200B, agent 100B
• Memory ID validation regex

[8.4.0] — 2026-06-23 — LIVING MEMORY

Added

• 7 lifecycle tools: memory_promote, memory_auto_promote, memory_decay, memory_consolidate, memory_link, memory_get_links, memory_build_links
• 4-phase memory lifecycle: promote → decay → consolidate → link graph
• Spreading activation via link graph (Collins & Loftus 1975)
• Ebbinghaus decay (5%/30d) for unused memories
• Immunological tier with Mahalanobis anomaly detection
• HTML dashboard reports with Chart.js (dark theme)
• ONBOARDING.md, INSTALL_FOR_AGENTS.md