superlocalmemory 3.5.3__tar.gz → 3.5.5__tar.gz

{superlocalmemory-3.5.3/src/superlocalmemory.egg-info → superlocalmemory-3.5.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: superlocalmemory
-Version: 3.5.3
+Version: 3.5.5
 Summary: Information-geometric agent memory with mathematical guarantees
 Author-email: Varun Pratap Bhardwaj <admin@superlocalmemory.com>
 License: AGPL-3.0-or-later
@@ -93,9 +93,9 @@ Dynamic: license-file
   <img src="https://superlocalmemory.com/assets/logo-mark.png" alt="SuperLocalMemory" width="200"/>
 </p>
 
-<h1 align="center">SuperLocalMemory V3.4</h1>
+<h1 align="center">SuperLocalMemory V3.5</h1>
 <p align="center"><strong>Every other AI forgets. Yours won't.</strong><br/><em>Infinite memory for Claude Code, Cursor, Windsurf, and any MCP-compatible AI client.</em></p>
-<p align="center"><code>v3.4.51 "Recency Intelligence"</code> — <strong>Session context is now time-aware.</strong><br>Stale memories from old projects no longer surface. Ebbinghaus decay + FSRS stability. One command: <code>pip install -U superlocalmemory && slm restart</code></p>
+<p align="center"><code>v3.5.0 "Scale-Ready + Context Injection v2"</code> — <strong>Your database auto-migrates. 6-channel recall in &lt;1s. Core Memory Block + explicit pinning. CozoDB + LanceDB on the recall path.</strong><br>No manual migrations. No data loss. One command: <code>pip install -U superlocalmemory && slm restart</code></p>
 <p align="center"><strong>Backed by 3 published research papers</strong> (arXiv preprints + Zenodo-archived) · <a href="https://arxiv.org/abs/2603.02240">arXiv:2603.02240</a> · <a href="https://arxiv.org/abs/2603.14588">arXiv:2603.14588</a> · <a href="https://arxiv.org/abs/2604.04514">arXiv:2604.04514</a></p>
 
 <p align="center">
@@ -274,29 +274,36 @@ slm warmup    # Pre-download embedding model (~500MB, optional)
 pip install superlocalmemory
 ```
 
-### Upgrading to v3.4.5 "Scale-Ready"
+### Upgrading to v3.5.0 "Scale-Ready — CozoDB + LanceDB"
 
-**Migration is automatic.** Upgrade your package, restart the daemon, and your database migrates silently.
+**Migration is automatic.** Upgrade the package, restart the daemon — CozoDB, LanceDB, and the vector store all self-migrate in the background.
 
 ```bash
-# pip users
 pip install -U superlocalmemory
 slm restart
-
-# npm users  
-npm update -g superlocalmemory
-slm restart
-
-# Verify migration
 slm doctor
 ```
 
-No manual commands. No data loss. Your database upgrades in-place with zero downtime. The daemon auto-detects the old version and applies the migration on first start.
-
-**New capabilities after upgrade:**
-- Tiered storage: memories auto-classified as active/warm/cold/archived
-- Graph pruning: redundant edges removed, queries stay fast at 1M+ connections
-- Optional: `pip install superlocalmemory[cozo,lancedb]` for graph + vector acceleration
+No manual commands. No data loss. Your database upgrades in-place. The daemon applies all migrations (including CozoDB graph, LanceDB vector, and the `pinned` column for Core Memory) on first start after upgrade.
+
+**What you get after upgrading to v3.5.0:**
+- **CozoDB on the recall path** — entity_graph channel routes through the CozoDB backend (auto-detected, no config needed). Millions of graph edges indexed and traversed in milliseconds.
+- **LanceDB vector backend** — embedding search falls through to LanceDB when available (auto-detected, no config needed). Handles millions of vectors.
+- **6-channel recall <1s** — BM25→FTS5 (20ms vs 11s), Hopfield ANN prefilter (0.4s vs 6s), temporal fast-parse (0.25s vs 2.6s). All surfaces (MCP/CLI/Dashboard) use the same daemon path.
+- **Core Memory Block** — always-injected pinned facts (auto-derived + explicit pin/unpin via the `core_memory` MCP tool). Pinned facts surface even when the query doesn't match.
+- **Context Injection v2** — unified formatter, token-budgeted injection (mode-aware 2K/4K/8K), edge-placement ordering, full-fidelity content (no more 200-char stubs).
+- **Score normalization** — all fusion scores mapped to [0, 1] via soft-sigmoid. Monotonic — rank order preserved.
+- **Vector store auto-backfill** — facts with embeddings missing from the vector store are indexed on daemon start. No more "only 1/3 of corpus searchable."
+
+### Release History
+
+| Version | Codename | Key Features |
+|---|---|---|
+| **v3.5.0** | Scale-Ready + Context Injection v2 | CozoDB/LanceDB migration, 6-channel recall <1s, Core Memory Block, BM25→FTS5, context injection v2, score normalization |
+| **v3.4.5** | Scale-Ready (foundation) | Tiered storage (active/warm/cold), graph pruning, BackendOrchestrator scaffolding, CozoDB + LanceDB init + migration code (read path wired in v3.5.0) |
+| **v3.4.51** | Recency Intelligence | Ebbinghaus decay + FSRS stability, age gate, session context time-awareness |
+| **v3.4.22** | Scale-Ready (scaling) | v2 ranking pipeline (LightGBM, bandit ensemble), Hopfield channel, 6-channel parallel execution |
+| **v3.3.x** | Foundation | BM25Plus, Fisher-Rao manifold, sqlite-vec, RRF fusion, cross-encoder rerank. 3 published papers (arXiv 2603.02240 / 2603.14588 / 2604.04514) |
 
 ### First Use
 

{superlocalmemory-3.5.3 → superlocalmemory-3.5.5}/README.md

@@ -2,9 +2,9 @@
   <img src="https://superlocalmemory.com/assets/logo-mark.png" alt="SuperLocalMemory" width="200"/>
 </p>
 
-<h1 align="center">SuperLocalMemory V3.4</h1>
+<h1 align="center">SuperLocalMemory V3.5</h1>
 <p align="center"><strong>Every other AI forgets. Yours won't.</strong><br/><em>Infinite memory for Claude Code, Cursor, Windsurf, and any MCP-compatible AI client.</em></p>
-<p align="center"><code>v3.4.51 "Recency Intelligence"</code> — <strong>Session context is now time-aware.</strong><br>Stale memories from old projects no longer surface. Ebbinghaus decay + FSRS stability. One command: <code>pip install -U superlocalmemory && slm restart</code></p>
+<p align="center"><code>v3.5.0 "Scale-Ready + Context Injection v2"</code> — <strong>Your database auto-migrates. 6-channel recall in &lt;1s. Core Memory Block + explicit pinning. CozoDB + LanceDB on the recall path.</strong><br>No manual migrations. No data loss. One command: <code>pip install -U superlocalmemory && slm restart</code></p>
 <p align="center"><strong>Backed by 3 published research papers</strong> (arXiv preprints + Zenodo-archived) · <a href="https://arxiv.org/abs/2603.02240">arXiv:2603.02240</a> · <a href="https://arxiv.org/abs/2603.14588">arXiv:2603.14588</a> · <a href="https://arxiv.org/abs/2604.04514">arXiv:2604.04514</a></p>
 
 <p align="center">
@@ -183,29 +183,36 @@ slm warmup    # Pre-download embedding model (~500MB, optional)
 pip install superlocalmemory
 ```
 
-### Upgrading to v3.4.5 "Scale-Ready"
+### Upgrading to v3.5.0 "Scale-Ready — CozoDB + LanceDB"
 
-**Migration is automatic.** Upgrade your package, restart the daemon, and your database migrates silently.
+**Migration is automatic.** Upgrade the package, restart the daemon — CozoDB, LanceDB, and the vector store all self-migrate in the background.
 
 ```bash
-# pip users
 pip install -U superlocalmemory
 slm restart
-
-# npm users  
-npm update -g superlocalmemory
-slm restart
-
-# Verify migration
 slm doctor
 ```
 
-No manual commands. No data loss. Your database upgrades in-place with zero downtime. The daemon auto-detects the old version and applies the migration on first start.
-
-**New capabilities after upgrade:**
-- Tiered storage: memories auto-classified as active/warm/cold/archived
-- Graph pruning: redundant edges removed, queries stay fast at 1M+ connections
-- Optional: `pip install superlocalmemory[cozo,lancedb]` for graph + vector acceleration
+No manual commands. No data loss. Your database upgrades in-place. The daemon applies all migrations (including CozoDB graph, LanceDB vector, and the `pinned` column for Core Memory) on first start after upgrade.
+
+**What you get after upgrading to v3.5.0:**
+- **CozoDB on the recall path** — entity_graph channel routes through the CozoDB backend (auto-detected, no config needed). Millions of graph edges indexed and traversed in milliseconds.
+- **LanceDB vector backend** — embedding search falls through to LanceDB when available (auto-detected, no config needed). Handles millions of vectors.
+- **6-channel recall <1s** — BM25→FTS5 (20ms vs 11s), Hopfield ANN prefilter (0.4s vs 6s), temporal fast-parse (0.25s vs 2.6s). All surfaces (MCP/CLI/Dashboard) use the same daemon path.
+- **Core Memory Block** — always-injected pinned facts (auto-derived + explicit pin/unpin via the `core_memory` MCP tool). Pinned facts surface even when the query doesn't match.
+- **Context Injection v2** — unified formatter, token-budgeted injection (mode-aware 2K/4K/8K), edge-placement ordering, full-fidelity content (no more 200-char stubs).
+- **Score normalization** — all fusion scores mapped to [0, 1] via soft-sigmoid. Monotonic — rank order preserved.
+- **Vector store auto-backfill** — facts with embeddings missing from the vector store are indexed on daemon start. No more "only 1/3 of corpus searchable."
+
+### Release History
+
+| Version | Codename | Key Features |
+|---|---|---|
+| **v3.5.0** | Scale-Ready + Context Injection v2 | CozoDB/LanceDB migration, 6-channel recall <1s, Core Memory Block, BM25→FTS5, context injection v2, score normalization |
+| **v3.4.5** | Scale-Ready (foundation) | Tiered storage (active/warm/cold), graph pruning, BackendOrchestrator scaffolding, CozoDB + LanceDB init + migration code (read path wired in v3.5.0) |
+| **v3.4.51** | Recency Intelligence | Ebbinghaus decay + FSRS stability, age gate, session context time-awareness |
+| **v3.4.22** | Scale-Ready (scaling) | v2 ranking pipeline (LightGBM, bandit ensemble), Hopfield channel, 6-channel parallel execution |
+| **v3.3.x** | Foundation | BM25Plus, Fisher-Rao manifold, sqlite-vec, RRF fusion, cross-encoder rerank. 3 published papers (arXiv 2603.02240 / 2603.14588 / 2604.04514) |
 
 ### First Use
 

{superlocalmemory-3.5.3 → superlocalmemory-3.5.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superlocalmemory"
-version = "3.5.3"
+version = "3.5.5"
 description = "Information-geometric agent memory with mathematical guarantees"
 readme = "README.md"
 license = {text = "AGPL-3.0-or-later"}

{superlocalmemory-3.5.3 → superlocalmemory-3.5.5}/src/superlocalmemory/core/engine.py

@@ -369,6 +369,91 @@ class MemoryEngine:
             vector_store=self._vector_store,
         )
 
+    def store_fast(self, content: str, metadata: dict[str, Any] | None = None) -> list[str]:
+        """v3.5.5 WRITE-THROUGH: synchronous verbatim insert for IMMEDIATE recall.
+
+        Full ``store()`` blocks 30-180s on LLM fact-extraction + Ollama embedding
+        + graph building. That created a recall window: a memory stored via the
+        async path sat in pending.db, unrecallable, until the background
+        materializer caught up. An agent storing a decision then immediately
+        recalling it (same session, or a parallel/next session) would miss it.
+
+        store_fast inserts a verbatim AtomicFact (+ memory row) synchronously.
+        The FTS5 ``atomic_facts_fts`` trigger auto-populates on INSERT, so the
+        memory is **keyword/BM25-recallable the instant this returns** (~ms, no
+        LLM, no embedding). Embedding + entities + graph are enriched async by
+        the materializer (which detects facts with NULL embedding).
+
+        Returns real fact_ids immediately. Quality gate rejects template junk.
+        """
+        self._require_full("store_fast")
+        self._ensure_init()
+        import re as _re
+        import uuid as _uuid
+        from datetime import datetime, timezone
+        from superlocalmemory.storage.models import (
+            AtomicFact, FactType, MemoryRecord,
+        )
+        if not content or not content.strip():
+            return []
+        try:
+            from superlocalmemory.core.injection import is_low_quality
+            if is_low_quality(content):
+                return []
+        except Exception:
+            pass
+        now = datetime.now(timezone.utc).isoformat()
+        record = MemoryRecord(
+            profile_id=self._profile_id, content=content,
+            session_date=now[:10], metadata=metadata or {},
+        )
+        self._db.store_memory(record)
+        # Lightweight regex entities (matches store_pipeline verbatim path) so
+        # the entity_graph channel has something to work with before enrichment.
+        ents = sorted(
+            {m.group(1) for m in _re.finditer(
+                r"\b([A-Z][a-z]+(?:\s[A-Z][a-z]+){0,3})\b", content)}
+            | {m.group(1) for m in _re.finditer(r"\b([A-Z]{2,})\b", content)}
+        )
+        # v3.5.5: compute the embedding SYNCHRONOUSLY. A single warm embed is
+        # ~22ms (the 30-180s of full store() was LLM fact-extraction + graph,
+        # NOT embedding). With the embedding present, the semantic channel
+        # scores this fact correctly so it ranks properly IMMEDIATELY — not
+        # just keyword-findable but top-ranked. Graph/entity enrichment stays
+        # async. Embed failure → fact still inserted (keyword-recallable).
+        emb = None
+        fmean = fvar = None
+        try:
+            emb = self._embedder.embed(content) if self._embedder else None
+            if emb:
+                fmean, fvar = self._embedder.compute_fisher_params(emb)
+        except Exception:
+            emb = None
+        fact = AtomicFact(
+            fact_id=_uuid.uuid4().hex[:16], memory_id=record.memory_id,
+            profile_id=self._profile_id, content=content,
+            fact_type=FactType.EPISODIC, entities=ents,
+            observation_date=now[:10], confidence=0.7, importance=0.5,
+            embedding=emb, fisher_mean=fmean, fisher_variance=fvar,
+            created_at=now,
+        )
+        self._db.store_fact(fact)  # FTS5 trigger → immediately BM25-recallable
+        # Upsert to vector store so the semantic channel finds it now.
+        try:
+            vs = getattr(self, "_vector_store", None)
+            if emb and vs and getattr(vs, "available", False):
+                vs.upsert(fact.fact_id, self._profile_id, emb)
+        except Exception:
+            pass
+        # Persist BM25 tokens too (covers the in-memory rank_bm25 fallback path).
+        try:
+            bm25 = getattr(self._retrieval_engine, "_bm25", None)
+            if bm25:
+                bm25.add(fact.fact_id, content, self._profile_id)
+        except Exception:
+            pass
+        return [fact.fact_id]
+
     # -- Recall operations --------------------------------------------------
 
     def recall(

{superlocalmemory-3.5.3 → superlocalmemory-3.5.5}/src/superlocalmemory/mcp/tools_core.py

@@ -110,27 +110,45 @@ def register_core_tools(server, get_engine: Callable) -> None:
         Extracts atomic facts, resolves entities, builds graph edges,
         and indexes for 4-channel retrieval.
         """
+        meta = {
+            "project": project,
+            "importance": importance,
+            "agent_id": agent_id,
+            "session_id": session_id,
+        }
+        # v3.5.5 WRITE-THROUGH: route through the daemon's /remember, which does
+        # a synchronous verbatim insert (memory is keyword/BM25-recallable the
+        # instant this returns) and enqueues async enrichment. This closes the
+        # recall window so a parallel/next agent finds memories saved seconds ago.
+        # Falls back to pending.db only if the daemon is unreachable.
         try:
-            # v3.4.32: Store-first pattern. Write to pending.db and return
-            # immediately. The daemon's pending-materializer thread drains
-            # the queue with recall priority, so concurrent MCP remembers
-            # no longer contend with /search on the shared embedder.
-            from superlocalmemory.cli.pending_store import store_pending
-
-            pending_id = store_pending(content, tags=tags, metadata={
-                "project": project,
-                "importance": importance,
-                "agent_id": agent_id,
-                "session_id": session_id,
-            })
+            from superlocalmemory.cli.daemon import daemon_request, is_daemon_running
+            if is_daemon_running():
+                resp = daemon_request("POST", "/remember", {
+                    "content": content, "tags": tags, "metadata": meta,
+                })
+                if resp and (resp.get("fact_ids") is not None or resp.get("ok")):
+                    fids = resp.get("fact_ids") or []
+                    return {
+                        "success": True,
+                        "fact_ids": fids or [f"pending:{resp.get('pending_id','')}"],
+                        "count": len(fids) if fids else 1,
+                        "pending": not fids,
+                        "message": "Stored (recallable now; enriching async).",
+                    }
+        except Exception as dexc:
+            logger.debug("MCP remember via daemon failed, pending fallback: %s", dexc)
 
+        try:
+            from superlocalmemory.cli.pending_store import store_pending
+            pending_id = store_pending(content, tags=tags, metadata=meta)
             return {
                 "success": True,
                 "fact_ids": [f"pending:{pending_id}"],
                 "count": 1,
                 "pending": True,
                 "pending_id": pending_id,
-                "message": "Stored — facts will appear in the dashboard shortly.",
+                "message": "Stored — facts will appear shortly (daemon offline).",
             }
         except Exception as exc:
             logger.exception("remember failed")

{superlocalmemory-3.5.3 → superlocalmemory-3.5.5}/src/superlocalmemory/server/unified_daemon.py

@@ -1582,14 +1582,28 @@ def _register_daemon_routes(application: FastAPI) -> None:
             extra = getattr(req, "metadata", None)
             if isinstance(extra, dict):
                 meta.update(extra)
+            # v3.5.5 WRITE-THROUGH: synchronous verbatim insert → the memory is
+            # keyword/BM25-recallable the instant this returns (~ms). Closes the
+            # recall window so a parallel/next agent finds memories saved seconds
+            # ago. Embedding/graph enrichment is deferred to the materializer.
+            fact_ids: list[str] = []
+            try:
+                fact_ids = engine.store_fast(req.content, metadata=meta)
+            except Exception as fexc:
+                logger.warning("store_fast failed, falling back to pending-only: %s", fexc)
+            # Enqueue for async enrichment (embedding + entities + graph). The
+            # materializer detects the already-inserted verbatim fact and enriches
+            # it in place rather than duplicating.
             pending_id = store_pending(
                 req.content, tags=req.tags or "", metadata=meta,
             )
             return {
                 "ok": True,
+                "fact_ids": fact_ids,
+                "count": len(fact_ids),
                 "pending_id": pending_id,
-                "status": "queued",
-                "note": "materialized async; pass ?wait=true for legacy sync",
+                "status": "stored" if fact_ids else "queued",
+                "note": "write-through: recallable now; enriching async",
             }
         except Exception as exc:
             raise HTTPException(500, detail=str(exc))
@@ -1816,14 +1830,36 @@ def _start_pending_materializer() -> None:
                     try:
                         import hashlib
                         content = item["content"]
-                        # Dedup: skip if identical content already stored.
                         content_hash = hashlib.md5(content.encode()).hexdigest()
+                        # v3.5.5: the write-through path already inserted a
+                        # verbatim fact (recallable via BM25). If it lacks an
+                        # embedding, ENRICH it in place (compute embedding +
+                        # upsert vector store) rather than skipping — otherwise
+                        # the fact would never be semantically searchable.
                         dup = engine._db.execute(
-                            "SELECT 1 FROM atomic_facts WHERE "
-                            "content = ? LIMIT 1",
+                            "SELECT fact_id, embedding FROM atomic_facts "
+                            "WHERE content = ? LIMIT 1",
                             (content,),
                         )
                         if dup:
+                            try:
+                                row = dict(dup[0])
+                                if not row.get("embedding") and engine._embedder:
+                                    emb = engine._embedder.embed(content)
+                                    if emb:
+                                        upd = {"embedding": emb}
+                                        try:
+                                            fm, fv = engine._embedder.compute_fisher_params(emb)
+                                            upd["fisher_mean"] = fm
+                                            upd["fisher_variance"] = fv
+                                        except Exception:
+                                            pass
+                                        engine._db.update_fact(row["fact_id"], upd)
+                                        vs = getattr(engine, "_vector_store", None)
+                                        if vs and getattr(vs, "available", False):
+                                            vs.upsert(row["fact_id"], engine._profile_id, emb)
+                            except Exception as eexc:
+                                logger.debug("enrichment of write-through fact failed: %s", eexc)
                             mark_done(item["id"])
                             continue
                         import json as _json

{superlocalmemory-3.5.3 → superlocalmemory-3.5.5/src/superlocalmemory.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: superlocalmemory
-Version: 3.5.3
+Version: 3.5.5
 Summary: Information-geometric agent memory with mathematical guarantees
 Author-email: Varun Pratap Bhardwaj <admin@superlocalmemory.com>
 License: AGPL-3.0-or-later
@@ -93,9 +93,9 @@ Dynamic: license-file
   <img src="https://superlocalmemory.com/assets/logo-mark.png" alt="SuperLocalMemory" width="200"/>
 </p>
 
-<h1 align="center">SuperLocalMemory V3.4</h1>
+<h1 align="center">SuperLocalMemory V3.5</h1>
 <p align="center"><strong>Every other AI forgets. Yours won't.</strong><br/><em>Infinite memory for Claude Code, Cursor, Windsurf, and any MCP-compatible AI client.</em></p>
-<p align="center"><code>v3.4.51 "Recency Intelligence"</code> — <strong>Session context is now time-aware.</strong><br>Stale memories from old projects no longer surface. Ebbinghaus decay + FSRS stability. One command: <code>pip install -U superlocalmemory && slm restart</code></p>
+<p align="center"><code>v3.5.0 "Scale-Ready + Context Injection v2"</code> — <strong>Your database auto-migrates. 6-channel recall in &lt;1s. Core Memory Block + explicit pinning. CozoDB + LanceDB on the recall path.</strong><br>No manual migrations. No data loss. One command: <code>pip install -U superlocalmemory && slm restart</code></p>
 <p align="center"><strong>Backed by 3 published research papers</strong> (arXiv preprints + Zenodo-archived) · <a href="https://arxiv.org/abs/2603.02240">arXiv:2603.02240</a> · <a href="https://arxiv.org/abs/2603.14588">arXiv:2603.14588</a> · <a href="https://arxiv.org/abs/2604.04514">arXiv:2604.04514</a></p>
 
 <p align="center">
@@ -274,29 +274,36 @@ slm warmup    # Pre-download embedding model (~500MB, optional)
 pip install superlocalmemory
 ```
 
-### Upgrading to v3.4.5 "Scale-Ready"
+### Upgrading to v3.5.0 "Scale-Ready — CozoDB + LanceDB"
 
-**Migration is automatic.** Upgrade your package, restart the daemon, and your database migrates silently.
+**Migration is automatic.** Upgrade the package, restart the daemon — CozoDB, LanceDB, and the vector store all self-migrate in the background.
 
 ```bash
-# pip users
 pip install -U superlocalmemory
 slm restart
-
-# npm users  
-npm update -g superlocalmemory
-slm restart
-
-# Verify migration
 slm doctor
 ```
 
-No manual commands. No data loss. Your database upgrades in-place with zero downtime. The daemon auto-detects the old version and applies the migration on first start.
-
-**New capabilities after upgrade:**
-- Tiered storage: memories auto-classified as active/warm/cold/archived
-- Graph pruning: redundant edges removed, queries stay fast at 1M+ connections
-- Optional: `pip install superlocalmemory[cozo,lancedb]` for graph + vector acceleration
+No manual commands. No data loss. Your database upgrades in-place. The daemon applies all migrations (including CozoDB graph, LanceDB vector, and the `pinned` column for Core Memory) on first start after upgrade.
+
+**What you get after upgrading to v3.5.0:**
+- **CozoDB on the recall path** — entity_graph channel routes through the CozoDB backend (auto-detected, no config needed). Millions of graph edges indexed and traversed in milliseconds.
+- **LanceDB vector backend** — embedding search falls through to LanceDB when available (auto-detected, no config needed). Handles millions of vectors.
+- **6-channel recall <1s** — BM25→FTS5 (20ms vs 11s), Hopfield ANN prefilter (0.4s vs 6s), temporal fast-parse (0.25s vs 2.6s). All surfaces (MCP/CLI/Dashboard) use the same daemon path.
+- **Core Memory Block** — always-injected pinned facts (auto-derived + explicit pin/unpin via the `core_memory` MCP tool). Pinned facts surface even when the query doesn't match.
+- **Context Injection v2** — unified formatter, token-budgeted injection (mode-aware 2K/4K/8K), edge-placement ordering, full-fidelity content (no more 200-char stubs).
+- **Score normalization** — all fusion scores mapped to [0, 1] via soft-sigmoid. Monotonic — rank order preserved.
+- **Vector store auto-backfill** — facts with embeddings missing from the vector store are indexed on daemon start. No more "only 1/3 of corpus searchable."
+
+### Release History
+
+| Version | Codename | Key Features |
+|---|---|---|
+| **v3.5.0** | Scale-Ready + Context Injection v2 | CozoDB/LanceDB migration, 6-channel recall <1s, Core Memory Block, BM25→FTS5, context injection v2, score normalization |
+| **v3.4.5** | Scale-Ready (foundation) | Tiered storage (active/warm/cold), graph pruning, BackendOrchestrator scaffolding, CozoDB + LanceDB init + migration code (read path wired in v3.5.0) |
+| **v3.4.51** | Recency Intelligence | Ebbinghaus decay + FSRS stability, age gate, session context time-awareness |
+| **v3.4.22** | Scale-Ready (scaling) | v2 ranking pipeline (LightGBM, bandit ensemble), Hopfield channel, 6-channel parallel execution |
+| **v3.3.x** | Foundation | BM25Plus, Fisher-Rao manifold, sqlite-vec, RRF fusion, cross-encoder rerank. 3 published papers (arXiv 2603.02240 / 2603.14588 / 2604.04514) |
 
 ### First Use