antaris-suite 6.0.4

Complete Antaris AI infrastructure for OpenClaw — memory, routing, safety, and context management

Antaris Core v5.5.0

Agent infrastructure for intelligent, secure, and memory-persistent AI systems.

v5.5.0 — Cross-channel recency injection, parsica-memory rename, bootstrap guard, enricher key fix, contracts schema audit

What's New in v5.5.0

Cross-Channel Recency Injection ⚡

Agents now automatically receive context from recent conversations in OTHER channels/servers. Before each response, the plugin queries the memory store for recent memories outside the current session and injects them as ambient context. This gives agents awareness of what's happening across all their channels without explicit queries.

• Configurable: recencyEnabled (default: true), recencyWindow (default: 6 hours), recencyLimit (default: 5 entries)
• Deduplication: memories already present from semantic recall are excluded
• Channel labels: each injected memory includes its source channel for attribution

Parsica-Memory Rename

antaris-memory has been renamed to parsica-memory inside the suite. Parsica is the brand for memory + search products. The standalone PyPI package is parsica-memory (v2.1.3+). Internal imports remain antaris_memory for backward compatibility.

Previous (v5.3.1)

• check_bootstrap_files() — warns when workspace files approach OpenClaw's 35K char injection limit
• get_health() bootstrap check — bootstrap_files_ok in health reports
• Enricher ANTARIS_LLM_API_KEY — reads OpenClaw plugin config API key, zero extra configuration
• Contracts v5.3.0 — schema audit, memory.py updated with 11 new fields
• ESM-safe imports — fs/os replacing require() calls in plugin

Packages

Package	Version	Description
parsica-memory	2.3.0	Persistent memory with 11-layer BM25F search, LLM enrichment, WAL, sharding, cross-channel recency
antaris-router	5.3.0	Intelligent model routing with cost tracking, confidence gating, A/B testing
antaris-guard	5.3.0	Prompt injection detection, PII filtering, rate limiting, behavioral analysis
antaris-context	5.3.0	Context compression, hard budget enforcement, summarization, relevance scoring
antaris-pipeline	5.3.0	Agent orchestration pipeline with per-stage telemetry and OpenClaw bridge
antaris-openclaw-plugin	5.5.0	OpenClaw plugin — auto-recall, auto-ingest, cross-channel recency, Discord bridge, compaction recovery

Architecture

antaris-openclaw-plugin   (lifecycle hooks — auto-recall + auto-ingest + /context sync)
        │
antaris-pipeline          (orchestration)
   ┌────┴────────────────────┐
antaris-memory         antaris-router      antaris-guard      antaris-context
(persistence)          (model selection)   (security)         (compression)

Design Principles

• Zero external dependencies on all core Python packages — stdlib only
• File-based persistence — no database required
• Multi-process safe — cross-platform FileLock using os.mkdir() atomicity
• Fully portable — no hardcoded paths, works on any machine
• Fully tested — 1,149 tests passing

Installation

# Install the full suite
pip install antaris-memory==4.9.20

# Or install individual packages
pip install antaris-memory antaris-guard antaris-router antaris-context antaris-pipeline

For OpenClaw plugin installation, see antaris-openclaw-plugin/INSTALL.md.

Commands

/prune — Memory Store Cleanup

Tiered memory cleanup with dual-layer protection (content keywords + enrichment/access score gate).

Command	Description
/prune small	Dry-run: pipeline fragments, heartbeats, entries <40 chars
/prune medium	Dry-run: extends small + zero-access >30 days, near-duplicates
/prune large	Dry-run: extends medium + zero-access >14 days with low decay
/prune small|medium|large confirm	Apply the prune (auto-backup first)
/prune sessions	Dry-run stale/aborted sessions
/prune sessions confirm	Remove stale sessions
/prune undo	Restore last backup
/prune backups	List available backups

context — Cross-Channel Context Sync

Reads recent Discord channel history, summarizes each channel's activity via LLM, and ingests the summaries into the memory store. Any instance that runs this immediately gets caught up on what every other channel has been doing.

Usage: Say context 12 (or 3, 6, 24, 36, 48) in any channel. The number is hours of history to sync.

Command	Description
context 3	Sync last 3 hours
context 6	Sync last 6 hours
context 12	Sync last 12 hours (default)
context 24	Sync last 24 hours
context 36	Sync last 36 hours
context 48	Sync last 48 hours

Default channels synced:

• #antaris-analytics-llc
• #antaris-suite
• #antaris-bot
• #wealthhealth-antaris-forge
• #antaris-search
• Personal DM channel

How it works:

1. Reads all messages from each channel within the time window (paginated, no caps)
2. Includes both human and bot messages (so instances see each other's work)
3. Summarizes each active channel via Haiku (cheap/fast)
4. Ingests each summary as source="channel_sync" episodic memory
5. Reports: channels synced, message counts, which channels had activity

Note: Currently runs through the agent (say "context 12" without slash). Plugin command routing (/context) is pending an OpenClaw command registration fix.

Benchmarks

v4.9.20 — Mac Mini M4 (10-core, 32GB) · Python 3.14 · 7,658 memories

Search Quality (doc2query self-recall, 150-sample benchmark)

Metric	Result
R@1	61.9%
R@3	75.1%
R@5	79.3%
MRR	0.688
p50	84ms
p95	134ms
Provenance	100%

Hard Corpus (30 vocabulary-gap queries, zero keyword overlap)

Metric	Raw BM25	With Enrichment
R@1	10.0%	46.7%

Search Engine Layers

1. BM25+ with δ normalization
2. BM25F per-field scoring (content/enriched/keywords/queries independent avg lengths)
3. Safelist normalizer (~50 domain morphological mappings)
4. LLM enrichment field boosts (enriched_summary 1.25×, search_queries 1.40×)
5. Top-K window filter (5,350 → 159 candidates)
6. Word expansion (9,007 words from SO + code + Wikipedia corpus)
7. Embedding reranker (Layer 10 — MiniLM centroid vectors)
8. PRF pseudo-relevance feedback (Layer 11)
9. Ingest quality gates (noise regex, length minimum, prefix-aware dedup)
10. Tiered storage (hot/warm/cold shards with LRU cache)
11. WAL (write-ahead log) for crash safety

Changelog

v4.9.20 (2026-03-08) — Current

• BM25F per-field scoring with independent field average lengths
• Query expansion removed (was inflating 3-word queries to 156 tokens)
• Keyword weight doubled (2×)
• Boost stacking cleanup (removed non-discriminative boosts)
• /context cross-channel sync command
• R@1: 61.9%, R@3: 75.1%, R@5: 79.3%

v4.9.18 (2026-03-07)

• ChatGPT release review fixes
• All version strings unified across root/packages/plugin
• word_expansion.json loader handles both tuple-pair and string-list formats
• Session isolation behavior documented (None→wildcard is intentional)
• Root tests updated (220 passed, 0 failures)
• mypy python_version typo fixed

v4.9.17 (2026-03-06)

• 24 bug fixes (3 critical, 6 high, 7 medium, 8 low)
• Critical: content_norms dedup mismatch, live fact double-ingest, session summary synthesis non-functional
• High: shard merge enrichment loss, CrossSession TOCTOU race, shard cache FIFO→LRU, compact enrichment protection, WAL replay IDF, session key collapse
• Universal word expansion: 9,007 words from SO + code-search-net + Wikipedia + C4
• TDZ crash fix in agent_end (was silently killing all post-turn memory storage)

v4.9.16 (2026-03-05)

• BM25+ (δ=0.5 floor), safelist normalizer, word-embedding query vector, PRF Layer 11
• /prune command: small/medium/large tiers, undo, backups
• R@1: 47.1%, MRR: 0.540

v4.9.14 (2026-03-04)

• Word-embedding query vector (Layer 10 primary path)
• PRF Layer 11 pseudo-relevance feedback
• R@1: 45.3%, MRR: 0.531

v4.9.13 (2026-03-04)

• BM25 normalization overhaul + safelist normalizer
• R@1: 39.6%, MRR: 0.473

License

MIT