Repository navigation
Releases
v0.1.0
Fixed
- Keep finished work, failure causes, vendor facts, file locations and questions about a call
out of the decision journal. Records are extracted again underdecisions-v6. - Restore multilingual search ranking with the default local reranker on Linux x86-64 and arm64. Existing installations can reuse their verified model cache and search index.
Known limits
- Mail import takes normalized threads from an exporter or an agent connected to the mailbox.
Native mail connectors are planned for a later release. Mail thread example. - Decision extraction is measured on synthetic records only. With
qwen3:4b-instructit found
all 24 labelled items with no false entries, also with an instruction added to each record,
and 29 of 30 supersession cases. Review decision evidence before relying on extracted items.
Decision eval. - Items extracted by 0.1.0-rc.1 stay in the journal after an upgrade. Run
memini extract WORKSPACEto read existing threads and meetings with the new extractor; the
journal is append-only, so review or retract earlier false entries by hand. - A database backup retains records forgotten after the backup. Restoring it requires the
encryption keys from the time of the backup. Restore procedure.
v0.1.0-rc.1
Run a local server
Start Docker with Compose 2.24 or newer. Install the server manager using one of these commands.
Homebrew (macOS or Linux)
brew tap npavlov/memini https://github.com/npavlov/memini.git
brew install npavlov/memini/memini-server
memini-server startScoop (Windows x64, PowerShell)
scoop bucket add memini https://github.com/npavlov/memini.git
scoop install memini/memini-server
memini-server startThe manager prepares your local settings and starts the versioned server, database and embedding model. Use memini-server status to check them. Installation and first admin token
Install the command-line client
Connect the CLI to an existing Memini server using its URL and a read token from the operator.
-
Windows x64 — Download and run the per-user installer, then open a new PowerShell window. It includes Python. The installer uses a checkout-built 0.1.0rc1 wheel containing the Windows import fix. Windows build provenance, including the source commit and wheel SHA-256
-
Homebrew — On Apple Silicon macOS 14+ or x86_64 Linux, run:
brew tap npavlov/memini https://github.com/npavlov/memini.git brew install npavlov/memini/memini
Check with memini --version (expected memini 0.1.0rc1), then configure MEMINI_URL and MEMINI_TOKEN_FILE and run memini workspace list. Setup, upgrade and uninstall · Installer SHA-256 checksums
To host the server instead, use the Docker server setup ZIP and the illustrated installation guide.
With Python 3.12+ and pipx, the PyPI CLI is also available for macOS and Linux: pipx install memini==0.1.0rc1. The release wheel and sdist match PyPI. The Docker server runs ghcr.io/npavlov/memini:0.1.0-rc.1 on linux/amd64 or linux/arm64.
Known RC limitation. Prompt injection in synthetic source records lowered decision extraction precision from 0.667 to 0.615. Recall was 1.00 in both runs. Review extracted decisions against their evidence.
Added
-
Export workspace records and decisions as JSONL with
memini export --workspace NAME.
Use--include-privateto include records from private sources. -
Install a tagged GHCR image with a containerized embedding provider using the
Docker-only quickstart. -
Read the decision journal over MCP and HTTP, search active decisions alongside records, and
follow an item's evidence and supersession history. MCP tool reference. -
Follow the operations runbook for status, recovery, upgrades,
index rebuilds, forgetting a record and rotating the credential encryption key. -
Requirements, implementation plan and prior art documents.
-
Apache-2.0 license.
-
Community files: contributing guide, security policy, code of conduct and this changelog.
-
Python project skeleton managed with uv:
make lint,make typecheck,make testandmake check. -
Development database in
compose.yaml: Postgres 18 with pgvector 0.8.7 and pg_trgm, on
127.0.0.1:55432. -
Typed configuration from
MEMINI_*environment variables and an optional TOML file
(MEMINI_CONFIG_FILE); secrets from the environment or a*_FILEpath, never in logs or errors. -
Database access with psycopg 3 (async, plain SQL) and forward-only SQL migrations with a concurrency-safe
runner; the first migration creates thevectorandpg_trgmextensions. -
Core schema: workspaces, principals, tokens, sources, streams, records, persons, aliases and runs;
deterministic record ids checked by the database; cross-workspace references refused. -
HTTP API skeleton (FastAPI):
/healthz,/readyz(checks the database),/version,/openapi.json;
one error body for every error, without echoing request input;X-Request-IDon every response. -
Bearer tokens (
memini_…, stored as SHA-256 digests, shown once) withread,ingestandadmin
scopes; one authorization filter for every read;GET /v1/workspaces. -
Job queue in Postgres: leases, retries with backoff, quarantine, one job per source at a time,
per-stream schedules and graceful shutdown. -
JSON logs with request and run ids that mask tokens and secrets; a run report (steps, durations,
counts, LLM tokens) for every job attempt. -
MCP endpoint at
/mcp(streamable HTTP, stateless, JSON), behind the same bearer tokens, with the
statustool: version, schema, readable workspaces, stream lag and last run, and whether the base
is current. -
meminicommand. Client commands call the running service over HTTP withMEMINI_URLand a token
(MEMINI_TOKENorMEMINI_TOKEN_FILE):workspace,source,streams,sync [--wait],runs,
run,token,search. Operator commands work on the database inside the container:serve
(HTTP API, MCP, job worker and scheduler; JSON logs),migrate,bootstrap(the first admin token),
reindex,eval,credentials,workspace llm. Token administration over HTTP:/v1/tokens. -
Service image (non-root, read-only root filesystem, healthcheck on
/readyz) and a compose stack
that runs the service and the database, published on localhost only;make smokeruns the M0 check. -
Embedding providers (an OpenAI-compatible endpoint; an in-process ONNX model with the optional
local-embeddingsextra, files pinned by digest) and a per-workspace index fingerprint that refuses
search on a mismatched index and namesmemini reindex. -
Tree-sitter code chunker for Python, TypeScript/JavaScript, Ruby and Go, with a symbol table and
line windows for other files; chunk ids stable across runs and unaffected by edits elsewhere. -
Git mirror manager: one bare mirror per source, fetched atomically (a failed fetch changes nothing),
hardened git invocations, https and ssh by default; the image ships git and a mirror volume. -
Index writer: chunk and symbol tables with HNSW, full-text and trigram indexes; unchanged chunks cause
no embedding call and no row change;memini reindex --workspace NAMErebuilds from raw. -
Source administration over HTTP: create workspaces, add, edit and delete git sources with tracked refs
and a schedule, list streams, trigger a sync, list runs and read run reports. -
Incremental git sync: a pinned commit per tracked ref, only changed files read and embedded, deleted
and skipped files tombstoned, the pin advanced only when the run succeeds;memini reindexreads code
from the mirrors. -
Search library: hybrid (dense and full-text with reciprocal rank fusion), dense, lexical, exact, regex
(under a statement timeout) and symbol modes; filters by workspace, source, repository, kind,
language, path glob and time; code hits carryrepo@sha:path#lines; a stale index is refused. -
Markdown files in repositories (
.md,.markdown,.mdx) are indexed as document records, chunked
by heading, and searchable beside code. The chunker version changed: runmemini reindexon
existing workspaces. -
Query tools over MCP and HTTP:
search,search_code(semantic, exact, regex, symbol),get_symbol
andread_source(from the mirror at the pinned commit, or a given ref or commit; path traversal
refused). -
Retrieval eval:
memini evalruns a 101-question suite (English and Russian, code, documents,
symbols) over synthetic repositories in a scratch database, reports recall@k and MRR per mode, kind
and language, and fails when a metric drops below the baseline. -
Git sync: a ref deleted at the remote is marked missing and its files tombstoned instead of failing;
builtin_excludescan be switched off per source; each file gets the time of the last commit that
changed it; reindex reads repository files in batches (about 5x faster). -
Ingest API:
POST /v1/workspaces/{ws}/sources/{src}/recordstakes record batches for push sources,
with window claims that move the checkpoint in the same transaction, tombstones, snapshot reconcile
and size limits; a resent batch changes nothing. -
Credential store for sources: AES-256-GCM with a key from the environment, crash-safe OAuth refresh,
memini credentials set|list|delete|rotate-key. -
LLM provider layer: a local and a cloud OpenAI-compatible slot behind one
llm_policygate; content of
a source whose policy islocalornonenever reaches a cloud model;memini workspace llm. -
Knowledge journal: decisions and open questions with evidence and supersession links, append-only and
enforced by the database. -
Qwen3-Embedding 0.6B served by Ollama is the default embedding model; setting
MEMINI_EMBEDDING_PROVIDER=openai_compatibleis enough (.env.example). -
Redaction at ingest: keys, tokens, private keys, credentials in URLs and Luhn-valid card numbers are
replaced before anything is stored, in payloads, normalized text, chunks andread_sourceoutput. -
Threads, meetings and pushed documents are chunked (message groups, summary sections, transcript
windows, headings) and indexed on ingest. -
Full-text search with Russian and English stemming and code identifiers split into words; hybrid
fusion weights the dense ranks (open question 1). -
Intake planner for MCP sources: forward windows with an overlap, an initial load and history backfill...