▄▄▄▄  ▗▞▀▚▖▄▄▄▄   ▄▄▄   ▄▄▄ ▄   ▄ ▄▄▄▄ ▄▄▄▄ 
█ █ █ ▐▛▀▀▘█ █ █ █   █ █    █   █    █ █  █ 
█   █ ▝▚▄▄▖█   █ ▀▄▄▄▀ █     ▀▀▀█ ▀▀▀█ ▀▀▀█ 
                            ▄   █ ▄▄▄█ ▄▄▄█ 
                             ▀▀▀            

One binary. One local SQLite file. One memory, shared by every MCP client on your machine.

memory39 is a Rust CLI + MCP server backed by SQLite + FTS5. Every MCP-capable AI tool on your machine (Claude Code, Claude Desktop, Codex, OpenCode, OpenClaw) reads and writes the same ~/.memory39/memory39.db, so a fact learned in one client is instantly recallable from any other. No cloud, no daemon, no sync.

Results are ranked by temporal-priority scoring: 0.4 x relevance + 0.3 x importance + 0.3 x recency (30-day half-life), so recent important matches surface first.

Why memory39

• Persistent across sessions: memories live in an on-disk SQLite file and survive restarts, CLI invocations, and MCP reconnects.
• One knowledge base across every MCP client: Claude Code, Claude Desktop, Codex, OpenCode, and OpenClaw all point at the same ~/.memory39/memory39.db. A fact stored from Claude is instantly recallable from Codex; a person stored via the CLI shows up in every MCP client. No syncing, no duplication.
• Local and private: no cloud, no account, no telemetry. Your memory is a single SQLite file on your machine that you can inspect, back up, or move by copying.
• Single binary, zero daemon: CLI for scripting (memory39 recall ...), MCP server on demand (memory39 mcp). Nothing runs in the background between calls.
• Portable DB path: point at a different database by exporting MEMORY39_DB=/path/to/other.db (supports ~/ expansion). Useful for project-scoped memory or isolated benchmarks.
• Cross-type discovery: the connect command links concepts across events, things, persons, and places in a single query, so relationships surface even when facts are stored as different memory types.

Performance

memory39 uses a bloom filter as a pre-check layer before FTS5 queries. On every recall, the bloom filter tests whether the query tokens exist anywhere in the database - if they definitely don't, the FTS5 query is skipped entirely, returning zero results in O(1) with no disk I/O.

Layer	When it runs	Cost
Bloom filter	Every recall query	~nanoseconds, in-memory
FTS5 search	Only if bloom says "maybe"	Full-text index scan

How it works:

• Unigrams + bigrams - every memory field is tokenized into individual words and adjacent-word pairs, all stored in the bloom filter. A query for "alice berlin" checks both tokens and the alice+berlin bigram.
• Unicode-normalized - tokens are lowercased with diacritics removed (matching FTS5's unicode61 remove_diacritics 2), so café and cafe hit the same entry.
• Prefix-safe - long tokens (>6 chars) that FTS5 would prefix-expand are never skipped, avoiding false negatives.
• Persisted - the bloom filter is saved to <db>.bloom alongside the database and loaded on startup. Rebuilt automatically after bulk writes.
• Zero false negatives - if a token exists in any memory, the bloom filter always says "maybe". It can only produce false positives (saying "maybe" when nothing matches), which just fall through to FTS5 as usual.

Configured for 600K items at 0.001% false positive rate.

Measured cost

Numbers from a personal DB (~300 memories, 144 KB SQLite file):

Query	Path	Avg per call	Throughput
Unknown fact	Bloom says "no match possible"; no disk I/O	~120 ns	~8.5M ops/sec
Known fact	Bloom says "maybe" -> FTS5 search runs	~245 us	~4K ops/sec

Negative queries are ~2000x faster than positive ones. "Nanoseconds, in-memory" is literal: a bloom check is a handful of hash probes into a bitmap sitting in L1/L2 cache.

Install

cargo install memory39

Install for any AI CLI / IDE

Installs the binary and auto-configures it for every MCP client detected: Claude Code, Claude Desktop, Codex, OpenCode, OpenClaw.

curl -fsSL https://raw.githubusercontent.com/alejandroqh/marketplace/main/h39.sh | sh

Single binary: CLI by default, MCP server with memory39 mcp.

Memory Types

Prefix	Type	What it stores
E#	Event (dated)	Something that happened/will happen, with date+time
U#	Event (undated)	Same, without a specific date
T#	Thing	Object, concept, or fact
P#	Person	Social memory about someone
L#	Place	Spatial memory about a location

Every memory has importance (0-10), emotion, and tags. The prefix + rowid (e.g. E3, T12) is the universal ID used by forget and alter.

---

CLI

Global Flags

Flag	Default	Description
--db <path>	memory39.db	SQLite database file
--ram	off	In-memory database (non-persistent)

Commands

event - Store an event

memory39 event "Had coffee with Alice" --date 2025-03-15 --people Alice --tags coffee,social

Arg/Flag	Required	Description
<event>	yes	What happened (max 255 chars)
--date	no	YYYY-MM-DD - omit for undated
--time	no	HH:MM (default 00:00)
--note	no	Additional note
--tags	no	Comma-separated tags
--importance	no	0-10 (default 5)
--emotion	no	positive, negative, neutral, or free text
--location	no	Where it happened
--people	no	Comma-separated names
--source	no	experienced, told, read, observed

thing - Store a fact or concept

memory39 thing "Rust edition 2024 requires rustc 1.85+" --category programming --confidence 9

Arg/Flag	Required	Description
<thing>	yes	What to remember (max 255 chars)
--desc	no	Description
--category	no	Free-text category
--tags	no	Comma-separated tags
--importance	no	0-10 (default 5)
--emotion	no	Emotional valence
--source	no	Where this knowledge came from
--confidence	no	Certainty 0-10 (default 5)
--related	no	Comma-separated related concepts

person - Store a social memory

memory39 person "Alice" --role "ML engineer" --relationship colleague --met-at "KubeCon 2024"

Arg/Flag	Required	Description
<name>	yes	Person's name
--role	no	Role or title
--relationship	no	friend, colleague, family, etc.
--contact	no	Email, phone, handle
--met-at	no	Where/when you met
--last-seen	no	Last interaction YYYY-MM-DD
--note	no	Additional note
--tags	no	Comma-separated tags
--importance	no	0-10 (default 5)
--emotion	no	Emotional valence

place - Store a spatial memory

memory39 place "Blue Bottle Coffee" --address "123 Main St, SF" --kind building --tags coffee,work

Arg/Flag	Required	Description
<name>	yes	Place name
--desc	no	Description
--address	no	Address or coordinates
--kind	no	city, building, room, outdoor, virtual, etc.
--note	no	Additional note
--tags	no	Comma-separated tags
--importance	no	0-10 (default 5)
--emotion	no	Emotional valence

recall - Search memories

memory39 recall "coffee" --limit 5 --min 3 --kind event
memory39 recall "*"  # list all

Arg/Flag	Required	Description
<query>	yes	FTS5 search query (* = all)
-l, --limit	no	Max results (default 10)
--min	no	Minimum importance 0-10
--from	no	Date start YYYY-MM-DD (events only)
--to	no	Date end YYYY-MM-DD (events only)
--kind	no	event, undated, events, thing, person, place
--source	no	experienced, told, read, observed
--offset	no	Skip first N results (default 0)

connect - Find connections between concepts

memory39 connect Alice Berlin meeting

Arg/Flag	Required	Description
<concepts...>	yes	2-3 concepts to connect
--min	no	Minimum importance 0-10
--timeout	no	Timeout in ms (default 2000)

Three-phase discovery: (1) direct - all concepts in one memory, (2) shared - concepts in separate memories linked by a common field, (3) bridge - one-hop connections through shared field values.

forget - Delete a memory

memory39 forget E3

alter - Modify a memory

memory39 alter T2 --text "Updated fact" --importance 8

Flag	Description
--text	New primary text
--note	New note
--tags	New tags
--importance	New importance 0-10
--emotion	New emotion
--location	New location (events only)
--people	New people (events only)
--source	New source
--date	New date (dated events only)
--time	New time (requires --date)

---

MCP Server

memory39 mcp starts the MCP server (STDIO transport). It runs purely against the local SQLite database: no network calls, no API keys, no external services.

Database path: ~/.memory39/memory39.db (auto-created). This path is shared across every MCP client on the machine, so configuring memory39 in Claude Code, Claude Desktop, Codex, OpenCode, and OpenClaw gives all of them the same memory. Override with MEMORY39_DB=/path/to/other.db for project-scoped or isolated databases.

Configuration

Add to your MCP client config:

{
  "mcpServers": {
    "memory39": {
      "command": "memory39",
      "args": ["mcp"]
    }
  }
}

Tools

Tool	Description	Required Params
recall	Search with temporal-priority scoring	query
event	Store an event (dated or undated)	event
thing	Store a fact, concept, or object	thing
person	Store a social memory	name
place	Store a spatial memory	name
forget	Delete a memory by ID	id
alter	Modify fields of an existing memory	id
connect	Find connections between 2-3 concepts	concepts

All optional parameters match their CLI counterparts. See the tool schemas for full details.

Scoring

Results from recall are ranked by composite score:

Component	Weight	Description
Relevance	0.4	FTS5 match quality
Importance	0.3	Memory importance (0-10)
Recency	0.3	Exponential decay, 30-day half-life

Memory IDs

All tools use the same universal ID system:

Prefix	Table
E	Dated events
U	Undated events
T	Things
P	Persons
L	Places

---

Environment

memory39 has no external dependencies at runtime: no network, no API keys, no .env file. The only environment variable it reads is an optional DB-path override.

Variable	Used by	Purpose
MEMORY39_DB	MCP server	Override DB path (supports leading ~/). Default: ~/.memory39/memory39.db. For the CLI, use the --db flag instead.

Built With

• TurboMCP - Rust MCP server framework

License

Apache-2.0