keephive 0.13.0

A knowledge sidecar for Claude Code

keephive

A knowledge sidecar for Claude Code. It captures what you learn, verifies it stays true, and surfaces it when relevant.

Claude Code forgets everything between sessions. keephive rides alongside it using hooks, an MCP server, and context injection to give it persistent, verified memory.

Install

uv tool install keephive
keephive setup

Requires uv. This installs from PyPI, registers the MCP server, and configures Claude Code hooks.

Via pip:

pip install keephive
keephive setup

From source:

git clone https://github.com/joryeugene/keephive.git
cd keephive && uv tool install . && keephive setup

Stay up to date

hive up                    # upgrade in place (recommended)
uv tool upgrade keephive   # manual alternative; run keephive setup after

Run keephive setup again after upgrading manually to sync hooks and the MCP server registration to the new binary path.

Quick start

hive                                               # status at a glance
hive r "FACT: Auth service uses JWT with RS256"   # remember something
hive v                                             # verify stale facts
hive go                                            # launch interactive session
hive todo                                          # open TODOs

After a few sessions, hive shows what your agent has learned:

$ hive
keephive v0.13.0
  ● hooks  ● mcp  ● data

  4 facts (4 ok) | 12 today | 8 yesterday | 2 guides | 48K

  1 open TODO(s):
    [today] Add rate limiting to the /upload endpoint.

  Today:
  ~ [10:42:15] FACT: Auth service uses JWT with RS256, tokens expire after 1h.
  ~ [10:38:01] DECISION: Chose Postgres over SQLite for multi-user support.
  ~ [09:15:44] INSIGHT: The retry logic in api_client.py silently swallows 429s.
  [09:12:30] DONE: Migrate user table to new schema.

  Active draft: slot 1 · "api testing todo list..." (47 words)  ->  hive nc

  hive go (session) | hive l (log) | hive rf (reflect) | hive help

How it works

keephive uses the three extension points Claude Code exposes:

1. Hooks fire on events (session start, conversation compact, user prompt). They capture insights and inject context without any agent action.
2. MCP server gives Claude Code native tool access (hive_remember, hive_recall, etc.) so the agent can read and write memory directly.
3. Context injection surfaces verified facts, behavioral rules, stale warnings, matching knowledge guides, and open TODOs at the start of every session via the SessionStart hook's additionalContext field.

The loop

  capture --> recall --> verify --> correct
     ^                                 |
     +---------------------------------+

• Capture: The PreCompact hook extracts FACT/DECISION/TODO/INSIGHT entries when conversations compact. It reads the full transcript, classifies insights via LLM, and writes them to today's daily log.
• Recall: Captured entries surface automatically at the next session start via context injection. hive rc searches all tiers directly; hive shows current state.
• Verify: Facts carry [verified:YYYY-MM-DD] timestamps. After 30 days (configurable), they are flagged stale. hive v checks them against the codebase with LLM analysis and tool access.
• Correct: Invalid facts get replaced with corrected versions. Valid facts get re-stamped. Uncertain facts get flagged for human review.

Memory tiers

Tier	Path	Purpose
Working memory	~/.claude/hive/working/memory.md	Core facts, loaded every session
Rules	~/.claude/hive/working/rules.md	Behavioral rules for the agent
Knowledge guides	~/.claude/hive/knowledge/guides/	Deep reference on specific topics
Daily logs	~/.claude/hive/daily/YYYY-MM-DD.md	Append-only session logs
Archive	~/.claude/hive/archive/	Old daily logs after gc

Hooks

Hook	Trigger	What it does
SessionStart	New session	Injects memory, rules, TODOs, stale warnings
PreCompact	Conversation compacts	Extracts insights from transcript, writes to daily log
PostToolUse	After Edit/Write	Periodic nudge to record decisions
UserPromptSubmit	User sends prompt	Periodic nudge to record decisions

Commands

Command	Short	What
Capture
hive remember "text"	hive r "text"	Save to daily log
hive t <text>		Quick-add a TODO
hive note	hive n	Multi-slot scratchpad ($EDITOR)
hive n todo	hive 4 todo	Extract action items from a note slot
Recall
hive status	hive / hive s	Status overview
hive recall <query>	hive rc <query>	Search all tiers
hive log [date]	hive l	View daily log; hive l summarize for LLM summary
hive todo	hive td	Open TODOs with ages
hive todo done <pat>		Mark TODO complete
hive knowledge	hive k	List/view knowledge guides
hive prompt	hive p	List/use prompt templates
hive ps		Active sessions, project activity, git state
hive session [mode]	hive go	Launch interactive session
hive standup	hive su	Standup summary with GitHub PR integration
hive stats	hive st	Usage statistics
Verify
hive verify	hive v	Check stale facts against codebase; auto-corrects
hive reflect	hive rf	Pattern scan across daily logs
hive audit	hive a	Quality Pulse: 3 perspectives + synthesis
Manage
hive mem [rm] <text>	hive m	Add/remove working memory facts
hive rule [rm] <text>		Add/remove behavioral rules
hive edit <target>	hive e	Edit memory, rules, todos, etc.
hive skill	hive sk	Manage skill plugins
Maintain
hive doctor	hive dr	Health check
hive gc	hive g	Archive old logs
hive setup		Register hooks and MCP server
hive update	hive up	Upgrade keephive in-place

Features in depth

Log Summarize

hive l summarize pipes today's log entries to claude-haiku and prints 3-5 bullet-point highlights. Useful after long sessions before compaction.

Smart Recall

hive rc <query> uses an SQLite FTS5 index over all daily logs and the archive for ranked full-text search. Run hive gc to rebuild the index. Falls back to grep if the index is absent.

Notes

hive n is a multi-slot scratchpad. Each slot persists across sessions, auto-copies to clipboard on save, and can be initialized from a prompt template (hive n <template>). Use hive n.2 or hive 2 to switch to slot 2 and open it in $EDITOR.

hive n todo (or hive 4 todo for slot 4) scans the active slot for action items. Both plain text lines and bullet points from a ## todo section become candidates; items over 120 characters are skipped as observations. For a single item, you get a yes/no prompt. For multiple items, your $EDITOR opens with the candidates — delete lines you don't want, save and quit to confirm.

hive 4 "text" appends text directly to slot 4 without opening an editor. Use bare-digit commands with a quoted string to capture quick notes mid-session.

Edit

hive e <target> opens files in $EDITOR. Targets: memory, rules, todo (with diff-on-save), CLAUDE.md, settings, daily log, notes. Run hive e with no arguments to see all targets.

Sessions

hive go launches an interactive Claude session with your full keephive context pre-loaded.

Command	What
hive go	General session with full memory and warnings
hive session todo	Walk through open TODOs one by one
hive session verify	Check stale facts against the codebase
hive session learn	Active recall quiz on recent decisions
hive session reflect	Pattern discovery from daily logs
hive session <prompt>	Load a custom prompt from knowledge/prompts/

Reflect

hive rf scans daily logs for recurring patterns across multiple days. When it finds a theme, hive rf draft <topic> generates a knowledge guide from the matching entries. This is how scattered daily notes become structured reference material.

Audit

hive a runs three parallel LLM analyses on your memory state (fact accuracy, data hygiene, strategic gaps), then synthesizes them into a quality score with actionable suggestions.

Standup

hive su generates a standup summary from recent daily log activity and optionally includes GitHub PR data.

Stats

hive st shows usage statistics with per-project breakdown, session streaks, and activity sparklines.

Prompts

hive p lists reusable prompt templates stored in knowledge/prompts/. Use them to start notes (hive n <template>) or launch custom sessions (hive session <template>).

MCP tools

All commands are also available as MCP tools for Claude Code to call directly:

hive_remember, hive_recall, hive_status, hive_todo, hive_todo_done, hive_knowledge, hive_knowledge_write, hive_prompt, hive_prompt_write, hive_mem, hive_rule, hive_log, hive_audit, hive_recurring, hive_stats, hive_fts_search, hive_standup, hive_ps

Configuration

Variable	Default	Description
HIVE_HOME	~/.claude/hive	Data directory
HIVE_STALE_DAYS	30	Days before a fact is flagged stale
HIVE_CAPTURE_BUDGET	4000	Characters to extract from transcripts
ANTHROPIC_API_KEY	(unset)	Enables LLM features inside Claude Code sessions. Never needed from a terminal.
NO_COLOR	(unset)	Disable terminal colors

LLM features

TL;DR: If you have Claude Code, you pay nothing extra. Ever.

keephive calls claude -p, which runs under your existing Claude Code subscription. ANTHROPIC_API_KEY is never checked from a terminal or a hook — you cannot accidentally bill yourself.

The API path exists for one specific case: running LLM commands (hive a, hive v, etc.) inside a Claude Code session rather than from a separate terminal. That is the only time ANTHROPIC_API_KEY is consulted.

Two tiers

Tier	When active	Cost
claude -p subprocess	Terminal or hooks (default, always)	Free — Claude Code subscription
Direct Anthropic API	Inside Claude Code + ANTHROPIC_API_KEY set	Paid — per-call billing

Hooks (PreCompact, etc.) run without the CLAUDECODE environment variable, so they always take the free subprocess path regardless of whether ANTHROPIC_API_KEY is present in your shell.

LLM-powered commands

Command	Model	When to use
hive a (audit)	3× haiku + 1× sonnet	Intentional quality check
hive v (verify)	sonnet + tools, multi-turn	Validating stale facts
hive rf analyze/draft	haiku	Pattern discovery
hive su (standup)	haiku	Daily standup generation
hive l summarize	haiku	End-of-session summary
hive dr (doctor)	haiku, optional	Duplicate TODO detection
PreCompact hook	haiku	Automatic on compaction — always free

hive a and hive v are the heavyweight operations. Use them intentionally.

Free commands (no LLM)

hive r, hive rc, hive s, hive todo, hive t, hive m, hive rule, hive e, hive n, hive k, hive p, hive st, hive l (without summarize), hive gc, hive sk, and all hooks except PreCompact.

Disable automatic LLM calls

Set HIVE_SKIP_LLM=1 to skip the PreCompact hook's extraction step. SessionStart never calls an LLM.

Development

uv run pytest                          # all tests
uv run pytest -m llm -v -o "addopts="  # LLM E2E tests (slow, real API calls)
uv run pytest -x                       # stop on first failure

See CLAUDE.md for architecture details.

License

MIT