cortext v1.1.10

Patch release for public no-storage recall and package artifacts.

Fixes:

• Retention::Ephemeral now forces an input boundary across public text, audio, image, media, and replay ingress paths, so no-storage queries still run retrieval and hydrate Context::retrieved_memory.
• Durable write permission remains off for ephemeral calls, so CLI/package recall does not store the query as a new memory.
• Adds a public memory-eval runner plus packet, Codex, Grok, and AGY judge adapters.

Packages:

• PyPI: augmem.cortext==1.1.10, with six cross-platform native libraries and verified q8_0 model bootstrap. The model downloads to the user cache on first use to stay under PyPI's 100 MB file limit.
• npm: @augmem/cortext@1.1.10, with six N-API prebuilds and bundled q8_0 model/vocab assets.
• CLI: Zig-built archives for Linux x86_64/aarch64, macOS x86_64/aarch64, and Windows x86_64/aarch64.

Validation:

• cmake -S . -B build -DCORTEXT_BUILD_TOOLS=ON -DCMAKE_BUILD_TYPE=Debug
• cmake --build build --target cortext_cli cortext_memory_eval_runner cortext_tests -j
• ctest --test-dir build -R cortext_tests --output-on-failure passed in 492.62s.
• CLI smoke: stored The garage door code is 8841., ephemeral recall "garage door code" retrieved it at rel 0.96, and memory/signal counts stayed 2/2 before and after recall.
• Memory-eval smoke: 6 public ephemeral queries, 66 aggregate retrieved candidates, zero empty retrieval packets. Packet answer accuracy 0.0; Codex/Grok/AGY each 0.2 over five scored queries.
• Final PyPI wheel import returned 1.1.10; final npm package load returned 1.1.10.

cortext v1.1.9

Patch release for package documentation.

• Python and npm README examples now show real OpenAI Chat Completions SDK calls.
• Retrieved Cortext memories are passed as developer context messages rather than assembled into an ad hoc prompt string.
• Package/native versions are bumped to 1.1.9.

Published to PyPI as augmem.cortext==1.1.9 and npm as @augmem/cortext@1.1.9.

cortext v1.1.8

Patch release for PyPI/npm registry packaging.

• PyPI and npm packages now ship cross-platform native libraries/addons without embedding the large AIST GGUF model.
• Python and Node wrappers resolve a bundled/local model when present, otherwise download and checksum-verify q8_0 into the user cache on first engine creation.
• Package READMEs document install, first-use model cache behavior, LLM usage, durable-write behavior, and troubleshooting.
• Python package builder now clears stale egg-info/build metadata before wheel creation.
• Includes Zig-built cortext_cli archives for Linux, macOS, and Windows x64/aarch64.

cortext v1.1.7

v1.1.7

Patch release for the model-asset cutover and package readiness.

Highlights

• Removes the stale public models_dir API from C++, C API, and bindings.
• Uses bundled package model assets by default for Python and Node installs.
• Builds cross-architecture Zig native libraries for Python and Node packages.
• Builds installable Zig CLI artifacts for Linux, macOS, and Windows on x64 and arm64.
• Moves tools from the examples switch to the tools build path.
• Refreshes Python and npm READMEs with install, quickstart, LLM usage, runtime asset, and troubleshooting guidance.

Verification

• ctest --test-dir build-codex-clean -R cortext_tests --output-on-failure --timeout 900
• Python wheel smoke: import augmem.cortext, create Cortext(':memory:'), embed_text(...) returns 1280-dim unit vector.
• npm tarball smoke: install @augmem/cortext, create new Cortext(':memory:'), embedText(...) returns 1280-dim unit vector.
• Native Linux CLI artifact smoke: cortext_cli --help.

Notes

The Python wheel and npm tarball include the default AIST GGUF asset, so they are large. If a registry rejects the upload due file-size policy, use the GitHub release artifact while the registry limit is adjusted.

Cortext v1.1.6

Cortext v1.1.6

Patch release for the installable public CLI.

Highlights:

• cortext_cli builds under CORTEXT_BUILD_TOOLS=ON, separate from examples.
• Zig installs cortext_cli with -Dcli=true for Linux, Windows, and macOS targets.
• Ephemeral recall and default-knob stale-fact supersession evals remain passing.

Assets:

• cortext-v1.1.6-linux-x86_64.tar.gz
• cortext-v1.1.6-windows-x86_64.zip
• cortext-v1.1.6-macos-x86_64.tar.gz
• cortext-v1.1.6-macos-aarch64.tar.gz
• SHA256SUMS.txt

The CLI binary is packaged without model assets; pass --models or set CORTEXT_MODELS_DIR to a directory containing the runtime models. macOS binaries are unsigned.

Verification summary:

• CMake cortext_cli and cortext_tests rebuilt successfully.
• Focused eval/regression probes passed for ephemeral recall, supersession edge writing, stale-memory demotion, and streaming pacing.
• Zig ReleaseFast CLI artifacts built for Linux x86_64, Windows x86_64, macOS x86_64, and macOS aarch64.

Checksums:

2180e651075352beeb503a19bf538aa779fe14b3438d61df9add508921323193  cortext-v1.1.6-linux-x86_64.tar.gz
3aee486e9c5719fd1982898172a6ce66e6a7bf5501664023f913b8741a7c013f  cortext-v1.1.6-macos-aarch64.tar.gz
bc1d7aa27c6c0fa4745cef21a74d848ae423a065c6e73fec949102342ed0c4f6  cortext-v1.1.6-macos-x86_64.tar.gz
0f87cbd6eeb77ec781717c4a022da43013b6b11dd67fa168e3b1cbc49082922f  cortext-v1.1.6-windows-x86_64.zip

Cortext v1.1.0

Cortext v1.1.0 Release Artifact

This file is the tracked release artifact for the v1.1.0 line. The v1.1.0
tag should point at the commit that contains this artifact.

Release State

• Version: 1.1.0
• Runtime: hard-cut production engine, limited to embedding-backed memory,
  durable graph associations, bounded retrieval, and shallow consolidation.
• Product surface: unchanged from v1.0. v1.1.0 is a hardening release, not a
  feature expansion.
• Removed research stack: decoder, provider registry, semantic extractor,
  summarizer, static label bank, fact layer, label-bucket graph, and
  mode-selected deep consolidation remain preserved in git but not shipped.
• Public API status: C++ facade, C ABI, and binding entry points remain the
  release surface for this tag.

Version Surfaces

The tracked package surfaces report 1.1.0:

• CMakeLists.txt
• bindings/python/pyproject.toml
• bindings/javascript/package.json
• bindings/dart/pubspec.yaml
• build.zig
• build.zig.zon

cortext_version() and every FFI package should report the same version.

Fix Train

v1.1.0 ships the release-hardening fixes completed after v1.0:

• Fixed graph-retrieval temporal decay overflow/stale-recency behavior and
  recorded the frozen replay A/B result.
• Restored bounded SQLite WAL behavior: default file-backed auto-checkpointing,
  SQLITE_CHECKPOINT_TRUNCATE for full checkpoints, and objstore blob garbage
  collection during eviction.
• Fixed GitHub Actions drift, split AIST model coverage into a dedicated cached
  job, and added ASan+UBSan CI coverage over the non-AIST suite.
• Validated media ingress before filters, encoders, or payload copies touch
  caller buffers.
• Made accumulator windows explicitly volatile staging state and stopped
  restoring or flushing legacy aggregate-only accumulator snapshots.
• Compiled ablation environment hooks out of default builds behind
  CORTEXT_EXPERIMENT_HOOKS=OFF while documenting the operational environment
  variables.
• Removed stale label-classifier/human-label tooling, fixed eval env-file
  discovery, and removed stale vendored submodule entries.
• Added direct tests for previously untested tail pipeline operations, made
  can-fail tests assert state, scrubbed external CORTEXT_* env vars during
  test startup, and switched temp DB naming to unique pid/time/counter paths.
• Updated the paper with negative experiment results, pattern-separation
  definitions, current consolidation framing, and tracked eval artifact copies.
• Added root NOTICE and third_party/VERSIONS inventory.
• Documented store/transaction single-owner semantics and one-writer-per-DB
  runtime ownership; schema migrations now take BEGIN IMMEDIATE before
  reading applied migration IDs.

Current Evidence

The v1.1.0 public evidence inherits the v1.0 hosted Meta Multi-Session Chat
eval, the 128k RAG ablation, and the post-optimization full replay verification.
The cited aggregate artifacts are tracked under docs/paper/artifacts/; full
protocol details are recorded in docs/paper/sections/9_experimental.qmd,
docs/paper/sections/11_optimization.qmd, and the generated manuscript at
docs/paper/_manuscript/index.md.

Additional v1.1.0 local verification recorded during the fix train:

• Full ctest --test-dir build-test -R cortext_tests --output-on-failure
  passed after the test-integrity work.
• Full ctest --test-dir build-test -R cortext_tests --output-on-failure
  passed after the store ownership and migration-lock work in 557.72 seconds.
• Focused AIST, sanitizer, media-validation, WAL, objstore-eviction, migration,
  paper-render, JSON/JSONL artifact, and Zig smoke checks passed at their
  respective fix points.

Build Gate

cmake -S . -B build/release \
  -DCMAKE_BUILD_TYPE=Release \
  -DCORTEXT_DISABLE_OPENTELEMETRY=ON
cmake --build build/release -j
./build/release/tests/cortext_tests '~[aist]' --reporter compact

Run AIST-tagged tests separately on a host with the model installed:

./build/release/tests/cortext_tests '[aist]' --reporter compact

FFI Gate

cmake --preset ffi-release-node
cmake --build --preset ffi-release-node --target cortext cortext_node -j
(cd bindings/go && go test)
(cd bindings/dart && dart pub get && dart analyze && dart test)
python3 -m py_compile bindings/python/cortext/__init__.py
node --check bindings/javascript/index.js bindings/wasm/cortext.js examples/web/main.js
node -e "const c=require('./bindings/javascript'); if (c.version() !== '1.1.0') process.exit(1)"

Browser WebAssembly Gate

./build-wasm.sh
test -s build-wasm/dist/wasm/cortext.js
test -s build-wasm/dist/wasm/cortext.wasm

CI Gate

The GitHub workflow must pass with optional OpenTelemetry disabled for
deterministic model-free CI where applicable:

• Ubuntu native release build and non-AIST tests.
• Arch Linux native release build and non-AIST tests.
• Dedicated cached AIST model job running [aist].
• Debug ASan+UBSan non-AIST job.
• Browser WebAssembly bundle build.
• Zig host and Linux cross-build smoke checks.

Documentation Gate

When algorithm behavior or experiment conclusions change, update
docs/paper/sections/ and regenerate:

QUARTO_DISABLE_GIT=1 QUARTO_DISABLE_GITHUB=1 quarto render docs/paper

For packaging/API changes, update the root README and all affected binding
READMEs before tagging.

Cortext 1.0.0

Cortext 1.0.0

First stable release of the local memory engine behind augmem.ai.

This release is about the supported surface: the C/C++ API, the C ABI used by language bindings, embedding-only entry points, reset/external-store hooks, browser WASM, and model-free CI. The old experiment-only LLM/provider paths are not part of the release surface.

Artifacts

• cortext-1.0.0-linux-x86_64-ffi.tar.gz - CMake-built Linux shared library plus installed headers.
• cortext-1.0.0-zig-linux-x86_64.tar.gz - Zig-built Linux shared library plus headers for Zig/FFI consumers.
• cortext-1.0.0-browser-wasm.tar.gz - browser bundle with cortext.js, cortext.wasm, and the web demo scaffold.
• cortext-1.0.0-sha256sums.txt - checksums for the uploaded bundles.

Build Provenance

• Source tag: v1.0.0
• Commit: 25df8fe4b5df903d01c3c4762b4116aba2afbdfa
• CI: https://github.com/augmem/cortext/actions/runs/27878101454
• Passed jobs: ubuntu-native, arch-native, wasm-browser, zig-smoke

Notes

AIST model assets are not bundled. Runtime tests or demos that need embeddings expect models/AIST-87M-GGUF or CORTEXT_AIST_MODEL_PATH.

Cortext v1.0

Cortext v1.0 Release Artifact

This file is the tracked release artifact for the v1.0 line. The v1.0 and
v1.0.0 tags should point at the commit that contains this artifact.

Release State

• Version: 1.0.0
• Runtime: hard-cut production engine, limited to embedding-backed memory,
  durable graph associations, bounded retrieval, and shallow consolidation.
• Removed research stack: decoder, provider registry, semantic extractor,
  summarizer, static label bank, fact layer, label-bucket graph, and
  mode-selected deep consolidation.
• Public API status: C++ facade, C ABI, and binding entry points remain the
  release surface for this tag.

Version Surfaces

The tracked package surfaces now report 1.0.0:

• CMakeLists.txt
• bindings/python/pyproject.toml
• bindings/javascript/package.json
• bindings/dart/pubspec.yaml
• build.zig
• build.zig.zon

cortext_version() and every FFI package should report the same version.

Current Evidence

The current v1.0 public evidence is the hosted Meta Multi-Session Chat eval on
the public nayohan/multi_session_chat mirror, plus the 128k RAG ablation and
post-optimization full replay verification.

Hosted Frontier Judge

• Artifact:
  eval_runs/msc_frontier_late_200dlg_gpt55_20260630T053427Z/judge_openai_gpt55_four_system_clean.json
• Replay: 9,130 text turns from 708 public MSC rows.
• Protocol: daily source-time consolidation at 02:00 UTC, 5,000-event warmup,
  500-event probe stride, 9 probes, 3 hosted gpt-5.5 blind judgments per
  probe, 27/27 judgments complete.
• Result: Cortext won 21/27, traditional chat+RAG won 0/27, full-history upper
  bound won 1/27, hosted compaction rollup won 5/27.
• Mean context: Cortext 998 tokens versus traditional chat+RAG 49,196 tokens,
  a 97.97% reduction with probe-bootstrap 95% CI [97.77%, 98.17%].
• Caveat: this supports token-reduction, noise-reduction, and blind-win claims;
  it is not a completed sufficiency-match claim.

128k RAG Ablation

• Artifact:
  eval_runs/msc_rag_ablation_128k_gpt55_20260630T_actual/judge_openai_gpt55_rag_ablation_128k.json
• Protocol: same 9 MSC probes, 3 hosted gpt-5.5 blind judgments per probe,
  128,000-token judge-context cap, 27/27 judgments complete.
• Systems: Cortext native, semantic vector RAG, lexical keyword RAG,
  rolling-window chat, 16k hybrid chat+vector RAG, hosted compaction rollup.
• Result: Cortext won 19/27, hosted compaction rollup won 8/27, and all four
  RAG-style packet variants won 0/27.
• Prompt cap: max estimated judge prompt 116,425 tokens under the 128,000-token
  cap.
• OpenAI usage: 1,871,994 prompt tokens and 39,001 completion tokens across 27
  judge requests.

Full MSC Latency Verification

• Artifact:
  build/graph_profile/full_msc_verify_rerun_20260630T142345Z/summary.json
• Replay: same 9,130-turn MSC slice, same default knobs, daily consolidation,
  5,000-event warmup, and 500-event probe stride.
• Completion: 9 probes, 714 consolidations, 10,099 memories, 9,364 long-term
  memories, 21 working-memory rows, and 17,719 associations.
• Behavior: recursive non-timing probe content matched the saved frontier
  artifact and the prior post-optimization replay.
• Latency: judged probe total latency was 50.4-80.2 ms, mean 63.5 ms;
  GraphRetrieve.total was 22.2-29.2 ms, mean 25.9 ms.

Full protocol details are recorded in docs/paper/sections/9_experimental.qmd,
docs/paper/sections/11_optimization.qmd, and the generated manuscript at
docs/paper/_manuscript/index.md.

Build Gate

cmake -S . -B build/release \
  -DCMAKE_BUILD_TYPE=Release \
  -DCORTEXT_DISABLE_OPENTELEMETRY=ON
cmake --build build/release -j
./build/release/tests/cortext_tests '~[aist]' --reporter compact

Run AIST-tagged tests separately on a host with the model installed:

./build/release/tests/cortext_tests '[aist]' --reporter compact

FFI Gate

cmake --preset ffi-release-node
cmake --build --preset ffi-release-node --target cortext cortext_node -j
(cd bindings/go && go test)
(cd bindings/dart && dart pub get && dart analyze && dart test)
python3 -m py_compile bindings/python/cortext/__init__.py
node --check bindings/javascript/index.js bindings/wasm/cortext.js examples/web/main.js
node -e "const c=require('./bindings/javascript'); if (c.version() !== '1.0.0') process.exit(1)"

Browser WebAssembly Gate

./build-wasm.sh
test -s build-wasm/dist/wasm/cortext.js
test -s build-wasm/dist/wasm/cortext.wasm

CI Gate

The GitHub workflow must pass with optional OpenTelemetry disabled for
deterministic model-free CI:

• Ubuntu native release build and non-AIST tests.
• Arch Linux native release build and non-AIST tests.
• Browser WebAssembly bundle build.
• Zig host and Linux cross-build smoke checks.

Documentation Gate

When algorithm behavior or experiment conclusions change, update
docs/paper/sections/ and regenerate:

QUARTO_DISABLE_GIT=1 QUARTO_DISABLE_GITHUB=1 quarto render docs/paper

For packaging/API changes, update the root README and all affected binding
READMEs before tagging.