MemNet

Pure in-memory working-memory graph for LLM "goldfish brains" — structured state an agent writes once and re-reads each turn because the model forgets.

MemNet is scratch memory for the current task, not a durable knowledge base or vector store. Sessions live in RAM, expire by TTL, and disappear on session close. There is no automatic disk state; persistence is explicit and user-controlled via session save / session load.

The problem it solves

LLMs lose track of entities, state, and rules between tool calls. MemNet gives the agent a small, typed, atomised knowledge graph it can:

• atomise state into many small @TAG: rows and @EDG links (one idea per row — the most important discipline),
• add facts, tasks, and relations once; update them when state changes,
• pull back only the live slice on the next turn (query warm --anchor),
• settle or expire missions so they stop polluting the prompt,
• keep hard limits and produce machine-readable warnings instead of silent bloat.

The pipe wire format is intentionally token-efficient: warm reads inject a connected subgraph, not JSON dumps or prose blobs.

Installation

pip install memnet-llm

From source (development):

pip install -e ".[dev]"

The CLI command is still memnet. PyPI package name is memnet-llm (the name memnet on PyPI is a different project — memristive neural networks).

Requires Python ≥ 3.11.

Quick start

You need two terminals. One runs the in-memory server that holds the graph; the other runs the CLI client.

Terminal 1 (server):

memnet serve
# prints: MEMNET_SERVE=127.0.0.1:18765

Terminal 2 (client):

# See the cheat sheet
memnet guide --loose

# Start a session with the bundled schema
memnet session open --map-file src/memnet/examples/schema.example.txt
# stderr also prints: MEMNET_SESSION=mn_xxxxxxxx
$env:MEMNET_SESSION = "mn_xxxxxxxx"

# Ingest a small world (LAW rules + world state + missions)
memnet add --file src/memnet/examples/workflow.example.txt

# Preferred read for the next LLM turn: only live mission state
memnet query warm --anchor PLR01

# When done
memnet session close $env:MEMNET_SESSION

Without memnet serve running, any stateful command returns @ERR: serve_required.

For one-off scripting or tests you can set MEMNET_TEST_INLINE=1 to run in-process (no server), but this is not the normal multi-turn agent mode.

LLM agents: read LLM-GUIDE.md (in this repo) for the full agent playbook, the goldfish loop, settlement pattern, and disciplines. It is written to be consumed by models.

Atomisation (required)

MemNet is a knowledge graph, not a document store. Atomisation — breaking state into many small nodes and explicit @EDG edges — is the discipline that makes query warm token-efficient.

Do	Don't
One fact / entity / task per @TAG: row	Paragraphs or merged facts in one field
Wire relations with @EDG	"A helps B and also C" in a single row
Short fields: ids, codes, keys, paths	Prose, markdown, full file contents
Batch many lines in one add --stdin	One giant row instead of many atoms

# Good — three atoms + one edge
memnet add --stdin @"
@TSK: T01|Clear warehouse|1|in_progress|persistent
@NPC: N03|helper|labour|1|0|0|active|persistent
@EDG: E01|N03|helps|T01|labour|persistent
"@

See LLM-GUIDE.md for the full agent playbook. Application notes (novel, SysML, MUD) show domain-specific tag maps — all use the same atomisation rule.

The goldfish loop (recommended pattern)

A typical agent turn:

1. add new rows or update existing ones — atomised @TAG: lines (batch via --stdin or --file is best).
2. query warm --anchor <focus> — returns only active (non-recyclable) rows, always includes LAW.
3. Paste the wire lines into the prompt, reason, decide on next adds/updates or a mission settle.
4. On mission complete: update the TSK (or equivalent) with both status=settled and recycle=delete_on_settle. Mission edges usually use delete_on_expire or delete_on_settle.
5. Optionally housekeep prune stale --apply to physically remove settled rows.
6. Next turn starts again at step 1 with a (usually new) anchor.

query context (cold) returns everything and emits @WRN: stale_in_store|… on stderr when recyclable rows exist. Prefer warm.

Wire format

Token-efficient pipe rows — one graph atom per line (pair with atomisation above):

@TAG: field|field|...

• One record = one idea — split compound state into more rows + @EDG, not longer fields.
• Pipe inside a value must be escaped: note\|extra (or \\| in some shells).
• Always quote the whole line in PowerShell or bash when it contains special characters.
• Reserved output tags: @SESSION, @ERR, @WRN, @STAT, @REL, @DEL.
• Errors and advisories go to stderr; data rows to stdout.

Example multi-line ingest (PowerShell):

memnet add --stdin @"
@NPC: N01|Shen Tiexin|female(12)|0|traditional|80|active|persistent
@EDG: E01|N01|seeks_help|PLR01|unlock|delete_on_expire
"@

memnet update --stdin @"
@NPC: N01|Shen Tiexin|female(12)|0|traditional|90|active|persistent
"@

See memnet guide, memnet examples map, and memnet tagmap fields --tag <TAG> for the current schema.

Commands

Run any command with --help for full flags.

Session lifecycle

Command	Purpose
session open --map-file	Create a new session (prints @SESSION: and MEMNET_SESSION=... on stderr)
session resume <id>	Attach to an existing session
session current	Show the id from $env:MEMNET_SESSION (or "none")
session list	List live sessions (id, expires, minutes left, last modified)
session save --file	Optional export of the current graph to a user-chosen snapshot file (wire format)
session load --file	Restore a snapshot into RAM (new session id by default; --keep-id to reuse)
session close <id>	Destroy the session (graph is gone)

Default TTL is 60 minutes (MEMNET_SESSION_TTL_MINUTES or --ttl).

Add, update & delete

• add [line] [--file PATH] [--stdin] [--dry-run] [--allow-new-relation] [--agent NAME] [--session]
  • Create new rows only. Fails with id_exists if the id is already in the graph.
• update [line] [--file PATH] [--stdin] [--dry-run] [--allow-new-relation] [--agent NAME] [--session]
  • Replace existing rows only. Fails with not_found if the id is missing (catches update typos).
• Batch via --stdin or --file is strongly preferred. --dry-run parses without mutating.
• delete --id ID

Query (graph reads)

• query warm --anchor ID [--depth N] [--max-rows M] — the normal read for agents. active_only is forced; anchor is required. LAW rows are always included.
• query context [--anchor] [--depth] [--max-rows] [--active-only] — cold/full view (use for audit; warns on stale rows).
• query neighbors <id> [--depth]
• query path <src> <dst>

Direct reads

• read list [--tag T] [--active-only] [--where field=value ...]
• read get --id ID [--tag T]

--where filters by field value (exact match). Repeat for AND. Use * or ? wildcards for glob match (e.g. --where name=*Tiexin*).

Housekeeping

• housekeep stats — @STAT rows + caps for rows/edges/relations/orphans/dangling/recyclable/modified.
• housekeep stale|orphans|dangling|recyclable — list the respective sets.
• housekeep prune stale|... --apply — actually delete (emits @DEL lines and a summary on stderr).

stale = recyclable + dangling + orphans.

Schema & relations

• tagmap fields [--tag T] / map fields — reference field lists.
• tagmap show / map show — current session's effective tag map.
• relations list — allowed EDG relation names for this session.
• New relations are rejected unless add or update uses --allow-new-relation (subject to max_relations).

Examples & server

• examples map|workflow|add <tag>|path
• serve [--host] [--port] — the in-memory graph host. Required for normal CLI use across processes.
• version, guide, guide --loose

Session model & optional snapshots

• One session open per task. Agents should resume rather than open duplicates.
• The graph is RAM only while the session is live.
• session save --file my.snap writes a plain-text wire-format snapshot that you own. session load --file my.snap brings it back into a (usually new) RAM session.
• On close or TTL expiry the session is dropped from the server; no server-side files remain.
• Snapshots are a user convenience, not MemNet's durability layer.

Housekeeping, warnings, and signals

On every stateful command the server emits advisory @WRN lines on stderr (capped per call):

• near_cap*, ttl_expiring
• stale_in_store, stale_dangling, stale_orphans, stale_graph
• mission_settled
• fanout_clamped, dangling_endpoint, etc.

@STAT lines report counts vs caps. @DEL lines are emitted after successful prune --apply or delete.

Environment variables

Variable	Effect
MEMNET_SESSION	Default session id when --session is omitted
MEMNET_SESSION_TTL_MINUTES	Default TTL for new sessions (1..1440)
MEMNET_AGENT	Default agent name stamped on written records
MEMNET_SERVE_HOST, MEMNET_SERVE_PORT	Bind address for memnet serve (client discovery is implicit via the same vars)
MEMNET_MAX_ROWS / MAX_LAW / MAX_RELATIONS / ...	Hard caps (see Caps in config)
MEMNET_TEST_INLINE	When set, CLI runs in-process (tests, one-off scripts). Not for normal agent use.

All caps have MEMNET_MAX_* names; see source for the full list and defaults.

Architecture (one minute)

• tagMap — merged fixed + user schema loaded at session open. Drives parsing and field order on output.
• memStore — in-memory nodes + directed edges (EDG). Write-order index, simple BFS for neighbours / paths / warm packs.
• LAW rows — special, usually exempt from orphan/dangling accounting and always surface in warm reads.
• recycle field — persistent (default for active world) vs delete_on_settle / delete_on_expire (missions). query warm hides the latter.
• Server holds a registry of live SessionEntrys. CLI is a stateless client that talks over localhost TCP (or in-process in tests).

No JSON on the wire for LLM consumption — only the @TAG: lines plus a handful of control records on stderr.

Application notes

Seven self-contained worked examples live under application-notes/ — each shows a complete MemNet pattern (schema, seed, 6-step loop, domain LAW rows). Ordered by typical adoption path, not release date.

#	Note	Pattern
1	llm-software-development.md	Multi-turn coding in Cursor — @MOD/@SYM/@TSK/@USR/@DEC, verified locators, v0.2.12 session_load/session_save retrospective; complements grep/LSP/git and Cursor codebase indexing
2	llm-daily-news.md	Batch RSS digest (~100 articles/day) — run-scoped working memory (120-min TTL), @KYWD hub salience, @CLU/@SYN layers, Python bridge via send_command
3	llm-tech-docs-decomposition.md	Instrument manual / SCPI remote mode — R&S RTO rev 29, 4 584 @CMD, procedure layers with precedes/requires, two driver turns; scripts/extract_rto_scpi.py
4	llm-sysml-v2-modeling.md	SysML v2 textual modeling — 6U CubeSat PDU, @PKG cross-file refs, allocations/ports/traceability from rows, runtime behaviour budgeting
5	llm-novel-writer.md	Interactive novel / RPG — 6-step read → context → user-input-as-data → analyse → update → loop; @LORE/@SCN/@STEP, chapter merge
6	llm-mud.md	Multiplayer text MUD — server-side world agent + client prose agents, Alice in Wonderland sample; tiered atomisation, scripts/load_test_mud.py
7	llm-build-on-memnet.md	Builder guide — author your own MCP server (FastMCP + run_memnet bridge, JSON envelope, LAW supplementation) and Cursor skill pack (SKILL.md frontmatter, references/ split, mcp.json registration); worked example = mcp-memnet skill + novel-mcp application split

Supplement: novel-initial-state.md — bootstrap rows for the novel writer MCP pipeline.

Development

.\scripts\dev.ps1          # setup / test / lint / fmt / cli
pytest

Tests run with MEMNET_TEST_INLINE=1 so they do not require a separate serve process.

MCP (optional)

Install the MCP adapter:

pip install memnet-llm[mcp]

Prerequisites: memnet serve running in another terminal (same as normal multi-command CLI use).

# Terminal 1
memnet serve

# Terminal 2 — stdio MCP server for Cursor / other hosts
memnet-mcp

Set MEMNET_SESSION to your open session id. Example Cursor MCP config:

{
  "mcpServers": {
    "memnet": {
      "command": "memnet-mcp",
      "env": {
        "MEMNET_SESSION": "mn_your_session_id"
      }
    }
  }
}

Tools (v1): serve_status, session_open, session_current, query_warm, add, update, read_get, housekeep_stats. Each returns a JSON envelope with stdout / stderr wire lines, exit_code, session_id, and errors[] (from @ERR: lines). session_open auto-seeds LAW01–LAW05 on every new session (prepended on each query_warm); optional seed_lines adds @CFG anchors and domain @LAW rows in the same call. See LLM-GUIDE.md.

Licence

MIT