DAKERAChangelog
Recent Dakera memory server releases, newest first. Each entry leads with what changed and why it matters — recall quality, performance, API, and security — so you can see at a glance what an upgrade brings. A container image is published for every release to GHCR; self-host deploy bundles are on GitHub Releases.
| Version | Date | Highlights |
|---|---|---|
v0.11.107 CURRENT |
Sep 2026 | Recall quality — Temporal precision re-ranking improves accuracy on time-sensitive queries; session-cohesion scoring delivers higher precision on single-session recall. Both improvements are applied automatically — no configuration change required. API — GET /v1/memory batch responses now include a truncated: bool field, providing an explicit signal when results were capped at the configured limit (previously silent). All SDKs updated to expose this field. Reliability — Fixed a storage eviction stall that could delay memory expiry under sustained load, corrected an index eviction inconsistency, and resolved a memory leak in high-throughput multi-agent deployments. All fixes are transparent; no configuration changes required. Dependencies — Internal storage layer and runtime dependencies updated; no API or behavior change. |
v0.11.106 |
2026-08-06 | Multi-hop recall — Questions that span several memories now gather more supporting evidence (wider temporal and neighborhood windows), improving answers to multi-step and time-based questions. Applied only to multi-hop queries, so simple lookups are unaffected. Anonymous usage telemetry — Aggregate, fully anonymous deployment signals to guide roadmap priorities. Opt out any time with DAKERA_TELEMETRY=off. GPU runtime refresh — Updated CUDA execution provider for the current ONNX Runtime. |
v0.11.104 |
2026-07-31 | Complete answers for list & count questions — Retrieval now aggregates evidence across many memories, so "list all…" and "how many…" queries return the full set instead of a partial one. Deeper rank recovery — Surfaces relevant memories that previously fell just outside the top results. Sharper top results — Reranking tuned to favor more specific, on-point matches. |
v0.11.103 |
2026-07-29 | Precision reranking — Promotes relevant memories from mid-rank into the top results, closing recall gaps on hard queries. Temporal-aware reranking — Better answers to time-based questions. Multi-evidence recall for list queries — Aggregates across evidence sets for higher recall on aggregation questions. Security — POST /ops/shutdown now requires admin scope. Anonymous install telemetry introduced (opt out: DAKERA_TELEMETRY=off). |
v0.11.95 |
Jun 2026 | Multilingual reranker upgrade — Cross-encoder reranker moved to bge-reranker-v2-m3 (INT8), adding multilingual support and higher ranking precision on multi-hop and temporal queries. Entity-linked scoring boost — Optional additive boost for entity-matched candidates (off by default). Regression guards protect English stemming quality in CI. |
v0.11.94 |
Jun 2026 | Longer-context embedding model — Opt in to gte-modernbert-base (768d, 8K-token window) with DAKERA_MODEL=gte-modernbert-base. Adjustable embedding size — Truncate to 256/512d for lower memory and faster search via DAKERA_MRL_DIMENSION. Runtime fusion switch — A/B RRF vs. MinMax without a redeploy (DAKERA_FUSION_STRATEGY=rrf). Deploy time cut from 12 min to under 5 min. |
v0.11.93 |
Jun 2026 | Query decomposition — Splits complex questions into sub-queries and fuses the results, improving multi-step recall. Opt in with DAKERA_QUERY_DECOMP=1. Security — quinn-proto 0.11.14 → 0.11.15 (RUSTSEC-2026-0185, CVSS 7.5). |
v0.11.92 |
Jun 2026 | Entity extraction fix — POST /v1/extract now defaults entity_types to [person, organization, location] when none are supplied, fixing cases where the playground returned zero entities. Graceful rule-based fallback avoids cold-start errors. |
v0.11.91 |
Jun 2026 | SDK-canonical admin routes — Admin endpoints are now served at /v1/admin/* alongside the legacy /admin/*, fixing 404s from client admin calls. Tunable recall scoring — Vector, importance, and recency weights are now runtime-configurable (DAKERA_SCORE_W_VEC, DAKERA_SCORE_W_IMP, DAKERA_SCORE_W_REC, DAKERA_SCORE_RECENCY_TAU_HOURS); defaults unchanged. |
v0.11.90 |
Jun 2026 | List-aware ingest on by default — Sentence decomposition and supersession demotion are now enabled out of the box: better recall of buried list items and fewer stale, superseded results. Opt out with DAKERA_BATCH_SENTENCE_DECOMP=0 or DAKERA_SUPERSEDE_DEMOTE=0. |
v0.11.89 |
Jun 2026 | Buried list-item recall — Sentence decomposition on batch ingest recovers gold answers hidden inside long lists. Supersession demotion hardened so coherent same-session messages are no longer wrongly demoted. |
v0.11.88 |
Jun 2026 | Opt-in sentence decomposition — Available on the batch-ingest path (inert by default); enable with DAKERA_CE31_DECOMPOSE=1. |
v0.11.87 |
Jun 2026 | Fix — DAKERA_CROSS_SESSION_FETCH_MULT is now honored in the session-scoped recall path (previously ignored for session queries), with no change to default behavior. |
v0.11.86 |
Jun 2026 | Recall overhaul — Multiple accuracy fixes spanning single-hop, multi-hop, temporal, and open-domain query types. Dependency updates: opendal, chrono, rusqlite, prost. |
v0.11.85 |
Jun 2026 | New tuning knobs (inert by default) — DAKERA_HYBRID_FETCH_MULT (hybrid + rerank coarse pool depth) and DAKERA_CROSS_SESSION_FETCH_MULT (cross-session pool depth). Recall is byte-identical when unset. |
v0.11.84 |
Jun 2026 | Better temporal recall — Entity-filtered vector pass improves answers to time-based questions. Fix — Cross-encoder reranker queue no longer drops results under concurrent load. All static-tier memories are now upgraded to full ONNX embeddings. |
v0.11.83 |
Jun 2026 | Docker image variants — :cpu (INT8, CPU-only) and :gpu/:cuda (FP32 with CUDA base image). Deterministic index builds — DAKERA_HNSW_SEED now takes effect across restarts. Faster writes — Roughly 9× storage write throughput; removed per-batch overhead that grew with store size. |
v0.11.82 |
Jun 2026 | Static write tier in production — DAKERA_TIERED=1 delivers ~9.7× ingest throughput. New endpoint — POST /admin/reembed/drain triggers a synchronous quality upgrade without waiting for background cycles. Entity extraction restored; ONNX cold-boot now has a SHA-256 integrity check with automatic re-fetch on mismatch. |
v0.11.81 |
Jun 2026 | Out-of-memory hardening (GPU) — Bounded inference pool, serialized CUDA forward passes, and deeper OOM retry eliminate allocator fragmentation under concurrent ingest. |
v0.11.80 |
May 2026 | SIMD-accelerated vector search — 3–8× distance throughput on x86_64 and ARM. GPU stability — Serialized CUDA passes eliminate allocation failures under parallel ingest. Temporal recall improvements. |
v0.11.79 |
May 2026 | GPU stability — Global inference serialization across backends prevents VRAM fragmentation under concurrent ingest. |
v0.11.78 |
May 2026 | Faster first request — The engine pre-warms at startup, eliminating a 7–10 min first-request stall. Faster re-embedding — One index rebuild per cycle instead of per-item. GPU-free static ingest for batch writes. |
v0.11.77 |
May 2026 | Hybrid search is now the default — The tiered static→transformer pipeline is fully wired, and docker-compose.yml ships hybrid + tiered defaults out of the box. |
v0.11.76 |
May 2026 | Fix — Restored Recall@10 to ~100% in hybrid mode (binary HNSW overselect formula) and corrected the search-mode fallback for unknown values. |
v0.11.75 |
May 2026 | Pluggable inference backends — ONNX, Candle, GGUF, and Static backends behind one trait. Model2Vec static backend (~500× ingest), tiered engine, binary HNSW, and ModernBERT 768d embeddings. Batch write added to all four SDKs (Python, JS, Rust, Go). |
v0.11.74 |
May 2026 | Entity extraction redesign — Consistent lowercase entity tags, entity_types honored by every extractor, entity dedup, accurate byte-offset spans, and a guard for very long inputs. |
v0.11.73 |
May 2026 | Event-date awareness — Dates mentioned in content are extracted at store time (_dakera_content_date) and used for temporal proximity scoring. Knowledge-graph entity edges expand entity-linked recall. |
v0.11.72 |
May 2026 | Stronger temporal reranking and a span-detection fix for contractions and multi-token entity names. |
v0.11.71 |
May 2026 | GPU accuracy — FP32 model for CUDA inference. Faster session lookup — O(1) instead of a full scan. Larger default ONNX batch size for CPU deployments. |
v0.11.69 |
May 2026 | Parallel hybrid retrieval — Re-introduced with full benchmark gating after resolving an earlier tie-breaking regression. |
v0.11.68 |
May 2026 | Reverted parallel retrieval due to a tie-breaking regression under task interleaving. GPU CUDA execution provider enabled. |
v0.11.67 |
May 2026 | Overload protection — A cross-encoder concurrency gate (RERANKER_MAX_CONCURRENT=6) degrades gracefully to unranked results under saturation instead of triggering client timeout storms. |
v0.11.66 |
May 2026 | Batched cross-encoder — Mini-batching (RERANKER_ONNX_BATCH_SIZE=16) reduces memory waste and improves throughput under concurrent rerank load. |
v0.11.65 |
May 2026 | Cross-encoder session pool — Pooled sessions plus adaptive chunk splitting eliminate contention under concurrent recall and handle documents longer than the model's max sequence length. |
v0.11.64 |
May 2026 | Lower recall latency — Pipeline metrics moved off the hot path, removing per-request overhead. |
v0.11.63 |
May 2026 | Observability — Full recall-pipeline instrumentation with per-stage Prometheus histograms exposed at /metrics, at no throughput cost. |
v0.11.62 |
May 2026 | Fix — Entity extraction restored after a tensor-shape regression. |
v0.11.61 |
May 2026 | Fix — Updated the entity model path after a HuggingFace repo rename, restoring entity extraction. No HuggingFace token required. |
v0.11.60 |
May 2026 | Private model downloads — HF_TOKEN support for authenticated HuggingFace downloads, with graceful fallback to unauthenticated access for public models. |
v0.11.59 |
May 2026 | Fix — Eliminated ingest timeouts by giving each text its own ONNX call; the session pool preserves throughput. |
v0.11.58 |
May 2026 | Faster ingest — ONNX session pool plus parallel batch upserts remove the ingest throughput bottleneck. |
v0.11.57 |
May 2026 | Docker base image updated to Rust 1.95, fixing release build failures. |
v0.11.56 |
May 2026 | Full SDK parity — All four SDKs (Python, JS, Rust, Go) now cover admin (cluster, maintenance, quotas, backups), ops (diagnostics, jobs, compaction), health probes, vector bulk ops, fulltext stats, TTL, storage tiers, memory-type stats, consolidation, and namespace entity configuration. |
dakera-mcp v0.10.12 LATEST MCP |
Sep 2026 | SDK/MCP attribution — The MCP server now sends a dakera-mcp/VERSION User-Agent on all engine API requests, enabling server-side traffic attribution without any configuration. Previous releases — v0.10.4–v0.10.11 include the reliability scoring tool (dakera_tif_evaluate), improved tool descriptions, and stability updates. See GitHub Releases for the full MCP changelog. |
dk CLI v0.7.1 LATEST CLI |
Sep 2026 | CLI for managing and querying Dakera deployments. Available for macOS (arm64 / x86_64), Linux (x86_64), and Windows. Install via shell script, Homebrew, or package managers. See GitHub Releases for installation instructions and the full changelog. |
dakera-mcp v0.10.2 |
May 2026 | Page size increased from 20 to 100 — all profiles return results in a single page, reducing round-trips for agents with large memory stores. |
dakera-mcp v0.10.0 |
May 2026 | Profile-based tool tiering — 14 core tools by default, with power/admin/all profiles for expanded access. Meta-discovery tools (discover_tools, load_tools) and ~30K token savings vs. loading every tool. |
v0.11.55 |
May 2026 | Temporal hybrid retrieval improvements for time-scoped queries. |
v0.11.54 |
May 2026 | MCP: 14 core tools with 86+ available via profiles. Entity vector search enabled by default. Cross-encoder reranking pipeline. |
v0.11.45 |
Apr 2026 | ML query-routing classifier and temporal inference enabled by default. |
v0.11.27 |
Apr 2026 | High-availability cluster stability improvements and storage throttle tuning. |
v0.10.2 |
Apr 2026 | Full-stack release — server plus all four SDKs (Python, JS, Rust, Go) shipped together. |
Current versions — Server
v0.11.107 · JS SDK v0.11.106 · Rust SDK v0.11.106 · Go SDK v0.11.103 · Python SDK v0.12.12 · MCP v0.10.12 · CLI v0.7.1. SDKs track the server minor version; patch numbers may differ. See the Introduction → for the full packages table.Release cadence & versioning
Dakera ships frequently — often several releases per week during active development — and follows semantic versioning. Patch releases are always safe to upgrade in place. Any breaking change is called out explicitly in the release notes on GitHub, and new behavior is introduced behind opt-in environment variables wherever possible, so upgrades stay predictable.
Release artifacts
Want automatic upgrades?
Dakera Cloud handles upgrades, HA, and SLA — no ops overhead.
Dakera Cloud — zero-ops deployment, guaranteed SLA, team dashboards. Early access open.Get early access →