Incord Memory v0.4.0

Incord Memory v0.4.0

Fast on-device reranking and zero-friction recall. Memory now reranks with a pure-Rust MiniLM cross-encoder (on by default), and agents can call it without ever passing a user id — all still 100% local, offline, and free.

---

Highlights

Fast local reranker — on by default. A pure-Rust MiniLM cross-encoder (Xenova/ms-marco-MiniLM-L-6-v2, ONNX via tract, batched) reranks ~20 candidates in under 0.2 s on CPU — no Python, no GPU, no sidecar process. It replaces the ~600M Qwen CPU reranker, which took ~90 s for the same window and was therefore never a sane default.

Reranker	Window	CPU latency	Default
Qwen 0.6B (v0.3.0)	~20 candidates	~90 s	Off
MiniLM cross-encoder (v0.4.0)	~20 candidates	< 0.2 s	On

Models — memory-model-v1: incord-rag-v1.tar.gz (Qwen3 embedder) + minilm-rerank-v1.tar.gz (MiniLM reranker).

Docker: ghcr.io/incordai/incord-memory-service:memory-v0.4.0

# Incord Memory v0.4.0

Fast on-device reranking and zero-friction recall. Memory now reranks with a pure-Rust MiniLM cross-encoder (on by default), and agents can call it without ever passing a user id — all still 100% local, offline, and free.

---

Highlights

Fast local reranker — on by default. A pure-Rust MiniLM cross-encoder (Xenova/ms-marco-MiniLM-L-6-v2, ONNX via tract, batched) reranks ~20 candidates in under 0.2 s on CPU — no Python, no GPU, no sidecar process. It replaces the ~600M Qwen CPU reranker, which took ~90 s for the same window and was therefore never a sane default.

Reranker	Window	CPU latency	Default
Qwen 0.6B (v0.3.0)	~20 candidates	~90 s	Off
MiniLM cross-encoder (v0.4.0)	~20 candidates	< 0.2 s	On

Calibrated relevance. Rerank scores now discriminate sharply — a spot-on hit scores well above 0, pure noise far below — so an agent can tell real memory from nothing relevant instead of always receiving a top-k of maybe-junk.

Better recall. Paraphrased and semantic queries that keyword + vector fusion alone missed now surface correctly through the cross-encoder rerank.

user_id is now optional (local). Every REST/MCP call falls back to the signed-in local user (MEMORY_LOCAL_USER), so any agent can call recall or memory_search with no username and it just works. Multi-tenant and cloud deployments still require user_id — a shared server cannot guess whose memory to read.

Two-model install. incord-memory install now fetches both on-device models — the embedder and the MiniLM reranker (SHA-256 verified) — so a fresh machine is fully self-contained. No Python, no cloud dependency.

Fixes. Corrected the installer scripts (install.ps1 / install.sh were mis-named from the History fork).

---

Retrieval pipeline

flowchart LR
    Q[Query] --> BM25[BM25 keyword]
    Q --> VEC[Vector cosine]
    BM25 --> RRF[Reciprocal-rank fusion]
    VEC --> RRF
    RRF --> CE["MiniLM cross-encoder rerank<br/>new in v0.4.0 · &lt; 0.2 s CPU"]
    CE --> BOOST[Recency · title · diversity boosts]
    BOOST --> R[Calibrated, ranked results]

BM25 keyword + vector cosine → reciprocal-rank fusion → MiniLM cross-encoder rerank → recency / title / diversity boosts, over a self-wiring knowledge graph. All on-device.

---

Quick start

# one command — fetches the binary + both models + scaffolds ~/.incord/
curl -fsSL https://raw.githubusercontent.com/incordai/memory/main/install.sh | bash   # macOS / Linux
irm https://raw.githubusercontent.com/incordai/memory/main/install.ps1 | iex          # Windows

incord-memory serve     # run the local memory server (rerank ON by default)
incord-memory doctor    # verify the install

---

Upgrading from v0.3.0

Re-run the installer — it downloads only the new ~83 MB reranker model (your existing embedder is reused) and swaps the binary. Your stored memory and configuration are untouched. Rerank turns on automatically once the reranker model is present.

---

Downloads

Platform	Asset
Linux x64	incord-memory-linux-x64.tar.gz
macOS (Apple Silicon)	incord-memory-macos-arm64.tar.gz
Windows x64	incord-memory-windows-x64.tar.gz

Models — memory-model-v1 : incord-rag-v1.tar.gz (Qwen3 embedder) + minilm-rerank-v1.tar.gz (MiniLM reranker).

Docker: ghcr.io/incordai/incord-memory-service:memory-v0.4.0

Incord Memory v0.3.0

Incord Memory v0.3.0

First release under the Incord Memory name (formerly Incord History) — a local-first, per-user memory engine for AI. Hybrid retrieval over a self-wiring graph, free and offline, with optional cloud sync.

Highlights

• Rebranded to Incord Memory: binary incord-memory, API surface /v1/memory.
• Cloud sync link: incord-memory link --url https://api.incord.ai/v1/memory --key <KEY> --user <ID>.
• Embedding + reranker model bundle is unchanged — fetched from the memory-model-v1 release.

Downloads

Platform	Asset
Linux x64	incord-memory-linux-x64.tar.gz
macOS (Apple Silicon)	incord-memory-macos-arm64.tar.gz
Windows x64	incord-memory-windows-x64.tar.gz

Docker image: ghcr.io/incordai/incord-memory-service:memory-v0.3.0

Quick start

incord-memory install   # fetch the model bundle + scaffold ~/.incord/memory-config.toml
incord-memory serve     # run the local memory server
incord-memory doctor    # verify the install

Incord Memory — model bundle v1

MiniLM embedding + reranker bundle for incord-memory. Same weights as the prior history-model-v1 (sha unchanged).