Neonmem 0.9.6

Neonmem 0.9.6

Imported knowledge that answers - even after you close and reopen the project.

Imports that stay answerable

• The importer's understanding now travels inside the cartridge. When you import docs, Neonmem trains a small model on your material so it can answer "what is X?" from it. That model now rides inside the single .neonmem file - so after you close and reopen a project, recall still finds the right answer, not just right after the import.
• No more dim-mismatch dead recall. Reopened, imported cartridges search with the same model their memories were built with, so semantic recall keeps working instead of silently falling back to noise.

Reliability

• Procedures & rules recall without erroring. Fetching a saved how-to or rule no longer hits an internal error.
• Memory survives an auto-compact. Your memory is saved before a session compacts and reloaded after - so a long session never loses its thread.
• Single, clean memory file. The cross-process write-lock now lives outside the cartridge folder (keyed by path), so your memory stays one tidy .neonmem file while the importer, the agent and the visualizer share it safely.

Platforms

• Windows, Linux (AppImage), with a Windows Store build in the pipeline. macOS on the way.

---

Local, private, and yours. Free for personal use.

Neonmem 0.9.5

Neonmem 0.9.5

A tougher importer that understands more - plus multi-process file safety.

Importer / Memory Builder

• Clear progress + a real "Done" signal. Big imports now show a progress bar and only say Done - memory saved when the cartridge is fully written and released.
• Can't be interrupted. The Builder window won't close mid-build, so an import can never be left half-written or collide with a read.
• Guidance & Max fields explained right in the UI.

Smarter imports (our own, fully offline)

• Cleaner memory. The loader now skips tables-of-contents, page numbers and headings, stops mislabelling reference text, and keeps the actual definitions.
• Imports code, not just docs. Classes, functions and interfaces become memories you can ask about by name.

Reliability

• Fixed a [WinError 183] .app.asar failure on some setups (all components now run from a safe working directory).
• Cartridges open with share-delete + retry, so reads never collide with a save.
• Cross-process single-writer lock so the importer, the agent and the visualizer can safely share one cartridge.

Platforms

• Linux (AppImage) available, alongside Windows. macOS is on the way.

---

Local, private, and yours. Free for personal use.

v0.9.4 - Reliable reopen after an update

v0.9.4 - Reliable reopen after an update

A stability fix: on some setups, after updating Neonmem and reopening an existing project, the first memory action could fail with a file error. Neonmem now always resolves your memory to a real, writable location, so projects reopen and just work across updates.

Includes everything from 0.9.3 (big memories load smoothly, even without a GPU).

Local, private, offline. Signed installer + portable. Windows & Linux.

v0.9.3 - Big memories, smooth even without a GPU

v0.9.3 - Big memories load smoothly, even without a GPU

A performance release for working with large memories (e.g. a whole folder of product docs imported into one project):

• Smooth big-import - building a large memory into a project no longer stutters the 3D view (renders the whole graph efficiently instead of one object per node).
• Runs without a dedicated GPU - the 3D brain now renders on demand (only when something changes) instead of continuously, so it stays responsive on corporate laptops, VDI and remote-desktop sessions with no hardware GPU.
• Point Neonmem at hundreds of docs and your agent becomes an expert on that subject, while the app stays smooth.

Local, private, offline. Signed installer + portable. Windows & Linux.

v0.9.2 - Importer that updates + sharper memory

v0.9.2 - An importer that updates, and a sharper memory

The importer now updates an existing cartridge in place. Re-run it over changed docs or code and it merges what's new, supersedes what moved on, and skips what it already knows - your memory stays current instead of just growing.

Also in this release (rolled up from 0.9.1):

• Resume from the exact stop-point - a fresh session picks up the open thread and your last directive, not just "recent notes".
• Connections render reliably - memories show their links (fixed an edge-dump issue), including across tiers.
• Cleaner subsystem names - the brain labels its self-organised subsystems with readable names, not raw code signatures.
• Version alignment - the CLI reports its real version (neonmem --version).

Offline, local, private. Signed installer + portable build. Windows x64.

v0.9.1 - Seamless restore

v0.9.1 - Seamless restore

A patch focused on resuming exactly where you left off.

• Live cursor - your last directive is captured the instant you submit it, so a fresh session resumes from the exact stop-point: both your prompt and the assistant's preceding action.
• Staleness guard - warns when memory may be behind recent work.
• Auto-capture - each finished turn's salient content (both sides) is encoded automatically.

All stored inside the binary memory cartridge - no sidecar files. Offline, local, private. Signed installer + portable build.

Neonmem 0.9.0 - the memory cartridge

Memory cartridge release.

Neonmem 0.8.2 - large-memory performance

Neonmem 0.8.2 - large-memory performance

A patch focused on big memories.

Fixed / improved

• Big graphs no longer crawl. Loading a large memory (tens of thousands of nodes and up, e.g. built with the file importer) used to drop the 3D view to a few FPS and fire animations for every node on load. Neonmem now switches to a large-graph mode above ~3,000 nodes: it skips the heavy per-node decoration and the per-frame idle animation, and wires edges in O(1) instead of O(N) - so load is far faster and the view stays responsive.
• Smaller memories look and behave exactly as before (full animation + glow).

Same memory core as 0.8.0/0.8.1.

A dedicated high-scale renderer (GPU point-cloud / instanced nodes) for 100k-800k-node memories is in progress - this release is the first, immediate improvement.

Downloads

• Neonmem-Setup-0.8.2.exe - Windows installer.
• neonmem-0.8.2-portable.tar.gz - portable build, no installer / no SmartScreen prompt. tar -xzf then run install.ps1. For locked-down / corporate PCs.

Neonmem 0.8.1 - import spinner fix

Neonmem 0.8.1 - bug-fix

A small patch over 0.8.0.

Fixed

• Memory Builder import spinner. When importing docs/code, the "Working." spinner could keep spinning after the import had already finished - the new nodes appeared in the visualizer, but the spinner never stopped. It now clears correctly when the import completes. (Cause was a CSS rule overriding the element's hidden state; the memory engine itself was never affected.)

No changes to the memory core - same engine as 0.8.0.

Downloads

• Neonmem-Setup-0.8.1.exe - Windows installer.
• neonmem-0.8.1-portable.tar.gz - portable build, no installer and no SmartScreen prompt. Extract with tar -xzf neonmem-0.8.1-portable.tar.gz, then run install.ps1. Best for locked-down / corporate PCs.

Neonmem 0.8.0 — gather facts from your docs & code

Neonmem 0.8.0

New: gather facts from your docs and code. Neonmem can now read your existing
documentation and source and keep the important facts as connected memory — so your
agent knows your APIs, config and code structure without re-reading the files.

What's new

• Fact ingestion — neonmem ingest-facts <path> (and the agent tool
  neonmem_ingest_facts) gather structured facts:
  • Docs (Markdown / text): API endpoints and what they do, config/data values,
    and field definitions — filler prose is skipped.
  • Code (Python AST): functions/methods + docstrings, a call graph
    (who-calls-whom), plus defined-in / part-of relationships.
  • Java (lightweight): classes, methods, dependencies (imports / extends), calls.
  • PDF / Word / images: Claude reads them with its own tools and hands the facts to
    Neonmem — no bulky parsers bundled in the app.
• Facts land in the same 3D graph and are recallable like any other memory. Secrets are
  auto-redacted and near-duplicates are skipped.

Also

• Includes the Windows import-hang stability fix (faiss/WMI).

Install

Download Neonmem-Setup-0.8.0.exe (Windows x64) and run it. Works with Claude Code
(full: session-start load + passive capture) and GitHub Copilot / VS Code
(experimental).

License

Free for personal, hobby, research, educational and nonprofit use under the PolyForm
Noncommercial License. Commercial use: sales@neonmem.com.