v0.1.0-alpha.32 — Keep history scans running past excluded paths

Git history containing protected paths could abort a scan with
PATH_OUT_OF_SCOPE: Select a relative path within this project. Collection now
skips those candidates and continues. Explicit out-of-scope queries remain rejected.
Renamed changes with a protected old path are also excluded before content capture.

Update and retry

codememory.cmd update --version v0.1.0-alpha.32

On macOS/Linux use codememory update --version v0.1.0-alpha.32.
Reconnect open Codex/Claude MCP clients. From the selected project directory:

codememory history scan --all
codememory history status

Use codememory.cmd on Windows. Existing captured history is retained; do not
remove the database, project registration or API profile. No new migration or
code reindex is required for this fix. Provider selection and credentials are unchanged.

Validation and limits

Local frozen installation, lint, typecheck, build and all 229 tests across 43 files
passed before tagging (full test duration: 91.70 seconds). All 15 package manifests
match the release version.

The regression reproduces the original exception without the fix and passes with
it. It uses a real temporary Git repository with protected files and a rename,
checks that excluded content is not captured, and keeps explicit path rejection.
Release publication requires the tagged Linux, Windows and macOS checks to pass.
The reported Windows log did not identify its triggering file; this fixes the
reproduced failure mechanism, not a claim of testing that private repository.

Installation ·
Changes

v0.1.0-alpha.31 — More focused project memory recall

Document-derived repository rules no longer all appear in every focused task.
Recall selects them using task words, document evidence paths or latest checkpoint
paths. Directly recorded repository rules remain eligible.

Behavior

• Exact file and directory-descendant evidence/checkpoint matches, with path component boundaries.
• English lexical matching with stemming; this is not semantic or cross-language retrieval.
• Existing approved content, scope and evidence links remain unchanged. Changed source does not silently revoke a rule.
• No-context recall remains a bounded overview. Use search_memory to browse all approved rules.
• Compatibility with databases that predate optional checkpoint/history tables.

Update

codememory.cmd update --version v0.1.0-alpha.31

On macOS/Linux use codememory update --version v0.1.0-alpha.31.
Reconnect Codex/Claude MCP after updating. No new migration or code reindex is
required specifically for this recall change.

Shared credentials, project permissions

The API profile and OS-stored credential are shared by projects under the same
local user account. Each project selects that profile and enables its workflows
separately. Updating preserves existing selections; do not recreate a profile or
re-enter its key for every project. A different device needs its own credential setup.

From a newly enabled project directory, with an existing deepseek profile:

codememory providers use deepseek --workflow both --yes
codememory history configure --documents --git --providers --yes
codememory memory configure --enable --yes
codememory history scan --all

Use codememory.cmd on Windows. Scanning collects evidence; selected evidence is
interpreted through the submit/worker/review flow. This release does not introduce
automatic approval or silently enable AI in every registered project.

Validation and limits

Publication requires the tagged Linux, macOS and Windows verification jobs to pass.
See the verification report for local results. Native credential stores and live
model quality on every host are not certified. This remains an alpha source release.

Setup ·
Verification ·
Changes

v0.1.0-alpha.30 — DeepSeek provider profiles and simpler CLI help

Use a shared API profile for selected project memory workflows, while keeping
collection local and permanent memory subject to evidence checks and approval.
Default CLI help now focuses on everyday work; advanced commands remain available.

Highlights

• DeepSeek adapter and an explicit OpenAI-compatible Chat Completions profile.
• Hidden key entry backed by the OS credential store; explicitly selected environment credentials for headless hosts. No plaintext key fallback.
• Independent history/conversation routing per project, shared per-profile usage allowances and bounded requests. Errors never silently switch providers.
• Selected provider/model information in workflow status, including separate verification calls to the same API model.
• Nine everyday root commands and five history commands in default help. Use --help-all to see every command and advanced option; existing scripts retain their command paths.
• Remove an unused legacy diff helper; typecheck now rejects unused locals and parameters.

Update

codememory.cmd update --version v0.1.0-alpha.30

On macOS/Linux use codememory update --version v0.1.0-alpha.30.
Reconnect Codex/Claude MCP after updating. The shared installation updates once;
provider selection and workflow permissions remain per project. No new database
migration or code reindex is required specifically for these changes.

Optional DeepSeek setup

codememory providers add deepseek --daily-usd 1 --daily-requests 100 --yes
codememory providers login deepseek
codememory providers test deepseek --yes

Use codememory.cmd on Windows. Login prompts for a hidden API key; do not put it
in chat or command arguments. The test makes one paid synthetic request without
project source. From the selected project directory:

codememory providers use deepseek --workflow history --yes
codememory history configure --documents --git --providers --yes
codememory history scan --all

Scanning collects evidence; the existing submit/worker/review flow interprets
selected excerpts. --workflow both also routes conversation memory; its separate
opt-in remains required. Creating a profile alone never enables transmission.

Full setup and operating limits

Verification and limits

Local lint, typecheck, all 226 tests across 42 files, build and macOS upgrade
checks passed before tagging. Publication requires Linux verification and Windows/macOS platform
checks, including new HTTP-provider and CLI-routing tests.

HTTP extraction/verification was tested using controlled responses and disposable
PostgreSQL databases. No real DeepSeek key or paid inference was used during this
implementation; real model quality and native credential-store availability on
all hosts are not certified. Linux key storage requires a working secret service.
Compatible providers may need an adapter for differing JSON/thinking behavior.

Daily dollar allowances are local estimates using configured rates, not a billing
guarantee. Unknown billing retains a reservation. The API verifier may use the
same model as the extractor in a separate call; this is not independent-model
agreement or human approval. This remains an alpha source release.

Verification report · Changes since alpha.29

v0.1.0-alpha.29 — Document evidence and Git change memory

CodeMemory can now collect Markdown evidence and actual file changes from the selected branch history. Low-information commit messages such as fx do not hide the changed files. Saved staged, unstaged and eligible untracked changes can be collected separately.

Highlights

• Project-scoped, resumable collection with progress, explicit gaps and exact Unicode evidence ranges.
• History CLI and MCP queries connect prior changes and documented intent to current source freshness.
• Spark/Haiku document candidates use source evidence; permanent rules require explicit approval. Collection alone sends no source to model providers.
• Existing PostgreSQL service, additive migration 7, indexed excerpt search, quotas, retention and scoped backup/restore.
• Matching English AGENTS.md and CLAUDE.md instructions.

Update and start

codememory.cmd update --version v0.1.0-alpha.29

On macOS/Linux use codememory update --version v0.1.0-alpha.29.
Reconnect Codex/Claude MCP after updating. The shared CLI update applies across registered projects; history settings are independently opt-in for each project.

From the registered project directory:

codememory history configure --documents --git --yes
codememory history scan --all
codememory history status

On Windows use codememory.cmd. Applying history configuration prepares migration 7. Keep PostgreSQL running. No code reindex is required solely to collect history. Model processing requires a separate history configure --providers --yes opt-in; automatic collection requires --automatic --yes and an open MCP connection.

Setup, evidence review and recovery

Validation and limits

Local validation passed 206 tests across 38 files, lint, typecheck, build and macOS upgrade checks before release preparation. The selected CodeMemory repository benchmark processed 42 commits and 1,109 file changes in 37.8 seconds; an unchanged repeat took 250 ms. These single-process measurements do not establish large-monorepo or concurrent-client performance.

The release workflow gates publication on Linux tests and macOS/Windows installer checks. New history provider prompts were tested with controlled providers, not live model inference. Syntactic locators do not constitute historical whole-program semantic analysis. Shallow history, exclusions and unavailable sources remain explicit coverage limitations. This is an alpha source release, not an npm package or standalone binary.

Verification report · Changes since alpha.28

v0.1.0-alpha.28 — Reliable evidence selection and clearer memory capture

Memory extraction now selects server-provided evidence IDs instead of rewriting
quotes. The server preserves original Turkish/Unicode text and message roles;
Haiku still verifies semantic support and permanent memory still requires approval.
Unknown or duplicate evidence IDs remain rejected. This removes model-authored
quote mismatch failures, not every possible INVALID_EVIDENCE or semantic error.

Highlights

• Protocol 2 keeps the public submission format and existing job history intact.
• Both AGENTS.md and CLAUDE.md define when to submit durable decisions, business
  rules and corrections, and when to skip routine commands or experiments.
• MCP status/tool guidance exposes capture triggers and pending-review behavior.
  These are agent instructions, not automatic transcript monitoring.
• Checkpoints share fingerprints within one request only. New requests recheck
  source; no persistent stale cache is introduced.

Update

codememory.cmd update --version v0.1.0-alpha.28

On macOS/Linux use codememory update --version v0.1.0-alpha.28.
Reconnect Codex/Claude MCP after updating to load the new runtime and instructions.
Run memory commands from the registered project directory or pass --project.
An existing failed job can be retried explicitly with codememory memory retry JOB_ID.
No new database migration is introduced beyond alpha.27.

Verification and limits

Local full tests, lint, typecheck, build and upgrade checks run before tagging.
A live synthetic Turkish Spark/Haiku flow completed extraction, approval and recall
after reconnect in 41.9 seconds. A warm-file synthetic comparison reduced reads
from 5 MiB to 1 MiB and median checkpoint time from 7.7 ms to 1.3 ms when five
memories shared one file. These are scenario measurements, not general latency
or Docker/PostgreSQL memory guarantees. See the repository verification report.

Whole-repository resource acceptance, automatic memory maintenance and symbol
rename repair remain open. Publication is gated on platform CI. This is an alpha
source release, not an npm package or standalone binary.

Changes since alpha.27

v0.1.0-alpha.27 — Claude schema compatibility and code checkpoints

This release corrects the inconsistent root package version that blocked alpha.26 publication. The alpha.26 tag remains unchanged.

Claude verification now receives JSON Schema Draft 7, addressing CLI rejection of
Draft 2020-12. Spark retains Draft 2020-12. Known schema rejections have a distinct
MEMORY_PROVIDER_SCHEMA_UNSUPPORTED diagnostic and exit status. Raw stderr is not
published because it may contain credentials or conversation text; unknown exits
no longer imply a login or quota problem.

Decision-to-code checkpoints

• Attach implementation, caller, Prisma model and test file paths to active
  memories through create_memory_checkpoint or the memory checkpoint CLI command.
• Inspect immutable history with get_memory_checkpoints or memory checkpoints.
• Recall checks current file hashes and flags changed, missing or unreadable
  evidence. Exact checkpoint path matches improve ranking among eligible memories.
• Add migration 6 and update English instructions in both AGENTS.md and CLAUDE.md.

Update

codememory update --version v0.1.0-alpha.27

Reconnect Codex/Claude MCP after updating. Normal managed setup applies database
migrations; external database installations must apply migrations before using
checkpoint writes. Run project memory commands in the registered project directory
or pass --project. Existing memory settings and history are preserved.

For an existing failed verification job, retry explicitly from its project:

codememory memory retry JOB_ID

This fix does not resolve INVALID_EVIDENCE quote mismatches. Do not retry jobs
indefinitely or bypass evidence checks.

Validation and limits

Local lint, typecheck, full tests and build pass before tagging. Tests include
provider-specific schema dialects, safe failure diagnostics, real MCP transport
with synthetic providers, checkpoint freshness and multi-project isolation.
Authenticated Claude CLI 2.1.267 was not retested locally for this release.
The release workflow separately gates publication on its platform checks.

Checkpoints observe file bytes, not behavior: symbol locators and test outcomes
remain caller reports. Automatic rename repair, whole-worktree attestation,
background memory maintenance and real kiosk performance acceptance remain open.
This is an alpha source release, not a standalone binary or an npm publication.

Installation ·
Changes since alpha.25

v0.1.0-alpha.25 — Evidence-backed project memory

Coding sessions can now retrieve relevant project rules at task start and submit
selected conversation evidence for review at task completion. The optional
workflow uses Spark to extract English candidates and Haiku to verify them against
original quotes. Nothing becomes active memory until explicitly approved.

What changes

• Add recall_context, workflow status, evidence submission, candidate inspection
  and explicit approval/rejection over MCP, with the same project routing as code
  queries.
• Add a durable PostgreSQL queue shared by client processes, bounded retries,
  classification correction, process cancellation on database lease loss and
  per-candidate approval provenance.
• Add codememory memory CLI commands. Both AGENTS.md and CLAUDE.md receive the
  English task-start/task-end protocol on project install and upgrade.
• Isolate model subprocesses from database credentials and customizations. Reject
  unexpected verifier models, fabricated quotes, incomplete reviews and schema
  drift. Support native and standard npm CLI layouts on Windows without passing
  evidence through a command shell.

Update and enable

codememory update --version v0.1.0-alpha.25

Reconnect the coding client's MCP connection after updating. In each project
where you want model-backed memory processing, after installing current Codex
and Claude CLIs and signing in:

codememory memory configure --enable --yes
codememory memory status

On Windows, codememory.cmd is also supported. The configure command applies
additive migration 5 and records opt-in for the selected project. No forced code
reindex is required. Model processing remains disabled elsewhere. Existing code
indexes and explicit memories are preserved.

Your connected MCP client runs the queue; codememory memory worker can run it
without a coding client. Inspect codememory memory candidates --job JOB_ID before
approving a record. See the memory guide
for approval, disablement, diagnostics and retry commands. New users can start
with the installation guide.

Validation and boundaries

Local lint, typecheck, build and all 186 tests passed. The synthetic live MCP test
completed in approximately 27.4 seconds: Spark extraction, Haiku verification,
explicit test approval, client reconnection and successful recall. The account
quota message was excluded. Haiku reported the requested model; Codex did not
report actual-model identity and this remains explicit in metrics. These timings
are one probe, not a performance guarantee. The release is published only after
its platform CI gates pass.

This release does not automatically read chat logs, guarantee correct inference,
resolve every semantic conflict, or maintain file/symbol freshness automatically.
Submitted evidence is retained locally until project removal. Inferred candidates
promote to repository-scoped rules; explicit notes retain existing finer scopes.
Model-provider quotas and data handling still apply. The stable production gates
remain open.

Compare changes

v0.1.0-alpha.24 — Use multiple projects in one AI conversation

CodeMemory can now serve several explicitly selected projects through one MCP connection, while keeping each project's index and memory separate.

What changed

• The installer enables session project attachment for Codex and Claude Code. The agent calls list_projects; registered roots reported by the client can be attached automatically. For an additional user-selected local project, the agent can call attach_project without editing MCP configuration.
• Dynamic attachment requires an already registered, enabled project. Unrelated registrations are not exposed. Every query and memory write must select a project when several are attached.
• Successful updates refresh the marker-managed instruction blocks in both AGENTS.md and CLAUDE.md for installations using --client both. Each gets complete instructions. User rules outside the markers are preserved, and legacy managed @AGENTS.md blocks are upgraded.
• Status requests above 20 diagnostic messages are reduced to 20 instead of failing input validation. The response reports requestedLimit, limitReduced and the normal continuation offset.
• Repeated --project arguments remain available for fixed multi-project connections.

Update

codememory update --version v0.1.0-alpha.24

On Windows, use codememory.cmd if needed. The updater refreshes existing enabled registrations according to their selected client; disabled and missing directories are skipped. Reconnect the MCP connection after the update so the new runtime, tools and instructions are loaded. No database migration or forced reindex is required.

Validation and limits

Local frozen installation, lint, typecheck, build and 169 tests across 29 files passed. Validation covers real MCP STDIO routing, client roots and explicit attachment, concurrent duplicate requests, rejected disabled/unregistered roots, cross-project symbol and memory isolation, watcher changes, instruction migration and diagnostic pagination. Publication is gated on Linux, macOS and Windows workflow checks.

Client root reporting varies. Where a client does not report the additional directory, automatic attachment relies on the agent following the installed instructions. Native Codex/Claude conversation behavior has not been separately verified. This does not make web-chat uploads local repositories or infer cross-project call graphs. Attached projects remain available until the connection closes. See session attachment boundaries.

Quick start · Changes since alpha.23

v0.1.0-alpha.23 — Control index scope with project ignore rules

Temporary workflow reports should not make your application index look broken. CodeMemory now excludes .workflow-tmp directories automatically at any depth in the selected project.

What changed

• Add a root .codememoryignore with gitignore-style patterns for project-specific scratch/output files. Changes participate in the index fingerprint and are reconciled on the next scan. Initial indexing and watcher pruning use these rules.

• The shared path policy excludes temporary reports before scanning and parsing, and also excludes them from watching and direct source reads.

• Syntax errors in these reports no longer cause incomplete: true. The exclusion remains visible in ordinary PROTECTED_PATH counts.

• Real-source syntax and binary-file diagnostics remain visible. Normal files such as src/report.ts are not excluded by their name.

Update

codememory update --version v0.1.0-alpha.23

Use codememory.cmd on Windows if needed. Reconnect the project's MCP connection and wait for indexing to finish. Normal reconciliation removes previously indexed temporary files from the graph without deleting files on disk. No database migration is needed.

Other custom scratch directories can be listed in .codememoryignore or the existing exclude list in .codememory/config.json; see indexing configuration.

Validation and limits

Local verification passed 166 tests across 29 files, lint, typecheck and build. Regression tests cover root/nested directory pruning, watcher filtering, direct-read exclusion and the distinction between temporary errors and actual source coverage failures. The release workflow gates publication on platform checks. This remains an alpha release with documented static analysis limits.

Quick start · Changes since alpha.22

v0.1.0-alpha.22 — Prisma block comments without lost models

Prisma schemas containing /** ... */ or /* ... */ can now be indexed without removing comments from the source. This fixes skipped models and the resulting unresolved relationships in other schema files.

What changed

• A string-aware adapter masks block comments only in the parser's memory copy. Source files and source previews retain the original text.
• LF/CRLF line endings, UTF-16 offsets and diagnostic line/column positions are preserved, including comments containing Unicode characters.
• Strings and line comments containing comment delimiters remain untouched. Unterminated block comments still report a syntax error.
• Regression coverage includes five models, seven related schema files, a static Prisma query, worker transport and persisted source previews.

Update and rebuild

codememory update --version v0.1.0-alpha.22

On Windows, codememory.cmd is also available. Reconnect the selected project's MCP connection afterward. Parser revision 8 triggers re-analysis even if source files have not changed; wait for indexing to complete and inspect codebase_status diagnostics. To start a CLI index explicitly, run codememory index in the project directory when no other index job owns its lock.

No database migration is introduced. Existing project registrations and original source files are preserved. Block-comment documentation is available in source previews; this adapter does not attach it as AST documentation metadata.

Verification and scope

Local verification passed 162 tests across 29 files. The release is gated by lint, typecheck, the full test suite, build and native Linux/macOS/Windows workflows. This is an alpha release; dynamic Prisma usage and other analysis limits still apply.

Quick start · Changes since alpha.21 · Report a regression