kvault

Persistent, structured memory for AI agents — plain Markdown, a CLI, zero services.

pip install knowledgevault

Your agent creates nodes (people, projects, notes), keeps every parent summary a rollup of what's below, and orients itself with one cheap command:

$ kvault tree
. « Knowledge Base » [3 children, 11 total] ~2026-06-07
  notes [1 children, 1 total] ~2026-04-11
    reading_list ~2026-04-11
  people [2 children, 5 total] ~2026-06-02
    contacts [2 children, 2 total] ~2026-06-02
      mike_torres ~2026-01-20
      sarah_chen ~2026-06-02
    friends [1 children, 1 total] ~2026-03-14
      alex_rivera ~2026-03-14
  projects [2 children, 2 total] ~2026-06-07
    launch_plan « Launch Plan — v2 » ~2026-06-07
    website_redesign ~2026-05-28

One outline line per node: title, size, and most-recent activity — about 15 tokens each, so a several-hundred-node KB orients an agent for a few thousand tokens. Anything pruned by --depth or --max-children is called out in place (…74 nodes below), so a partial view can never silently hide content.

Built for developers using AI coding tools who want their agent to remember things between sessions — contacts, projects, meeting notes, research — in a structured, navigable format. kvault needs no API keys, no hosted service, no database: any agent that can run shell commands can use it.

How it works

• A node is a directory containing a single _summary.md — YAML frontmatter plus Markdown. Leaf nodes are entities (a person, a project); parent nodes summarize their descendants.
• Parent summaries are the index. Every level is a comprehensive rollup of the subtree below it, written by the agent itself. Navigation is top-down reading, not blind grepping.
• Writes propagate. kvault write returns the full ancestor chain so the agent rewrites those summaries in one follow-up call — the "2-call write workflow."
• The KB instructs the agent. kvault init generates an AGENTS.md with the workflow, the rules (search before create, never fabricate, propagate everything), and a periodic maintenance playbook.

Quickstart (30 seconds)

pip install knowledgevault
kvault init ./my_kb --name "Your Name"

Then tell your agent:

"Use kvault CLI commands to manage my knowledge base at ./my_kb"

The agent reads the generated AGENTS.md and starts working.

Agent skill included. skills/kvault/SKILL.md carries the full workflow in the portable SKILL.md agent-skills format, so the agent loads it on demand from any directory — no per-KB setup. Install it wherever your tool discovers skills:

# Claude Code
cp -r skills/kvault ~/.claude/skills/kvault

# OpenClaw (per workspace)
cp -r skills/kvault ~/.openclaw/workspace/skills/kvault

# Other agents: copy into your tool's skills directory, or paste the
# SKILL.md body into its custom instructions

Already have data? Point your agent at an export from a chat, email, or notes tool — see docs/importing-data.md.

The 2-call write workflow

# Call 1: write the node (stdin = frontmatter + markdown body)
kvault write people/contacts/sarah_chen --create --reasoning "Met at NeurIPS" --json --kb-root ./my_kb <<'EOF'
---
source: manual
aliases: [Sarah Chen, sarah@example.com]
---
# Sarah Chen
Research scientist at Acme AI...
EOF
# → {"success": true, "changed": true, "did": "created people/contacts/sarah_chen",
#    "notes": [{"code": "autofilled", "text": "name=Sarah Chen", ...}],
#    "ancestor_paths": ["people/contacts", "people", "."],
#    "ancestors": [{path, current_content}, ...], "journal_logged": true}

# Call 2: the agent rewrites the returned ancestors, including root
kvault update-summaries --json --kb-root ./my_kb <<'EOF'
[
  {"path": "people/contacts", "content": "# Contacts\n...updated..."},
  {"path": "people", "content": "# People\n...updated..."},
  {"path": ".", "content": "# Knowledge Base\n...updated..."}
]
EOF

In human mode the same write narrates its decisions under the receipt:

Created: people/contacts/sarah_chen
  autofilled  name=Sarah Chen
Journal: journal/2026-08/log.md
Ancestors to update: 3  (people/contacts, people, .)

Re-sending identical content is a detected no-op — the file is not rewritten, mtime and created/updated stay put, so the recency signal in the tree stays honest:

Unchanged: people/contacts/sarah_chen
  unchanged   body and metadata identical — file not rewritten, created 2026-08-10, updated 2026-08-10 preserved
Journal: journal/2026-08/log.md

Required frontmatter: source, aliases — kvault stamps created/updated automatically.

What kvault tells you

kvault reports what it decided, not what you asked for. A note is emitted only when kvault invented a value, deliberately changed nothing, half-failed, hid something, or fell back — silence means the operation went exactly as asked. Notes render as indented lines under the receipt (human mode) and as a notes array in --json and over MCP, each {code, text, level} — a note's why/next ride in the JSON at every tier, and print in human mode at --explain. Batch commands collapse repeated notes by code ({code, count, examples}), so a 40-ancestor maintenance run emits one line, not forty. The vocabulary is closed — 11 codes:

Tiers. -q/--quiet (receipt and warnings only) → normal → --explain (adds each note's why and the exact next command) → --trace (adds lock waits and mechanics). Flags work before or after the subcommand; KVAULT_VERBOSITY=quiet|normal|explain|trace sets the tier for hooks and cron jobs (flags win; a typo silently means normal). partial notes survive even --quiet — silencing a half-failure on request is a footgun.

--strict exits 3 when any warning-class note (partial, skipped, a broken lock) was emitted — for CI and unattended runs. check rejects it: its exit codes are already a contract, and its human output is frozen.

Durable ops log. Every successful mutating command (CLI and MCP) appends one row to .kvault/logs.db: op, path, did, notes, changed/partial flags, duration, session. kvault log tail shows what this KB's other agents and sessions did recently; KVAULT_SESSION groups the commands of one logical task, KVAULT_OPS_LOG=0 disables. A failed append can never fail a write — the CLI surfaces the miss as a skipped note.

Full 0.13.0 detail — every new JSON field, per-command changes, frozen surfaces — is in the CHANGELOG.

Structure that holds up under many agents

A KB rots in a specific way when several agents write to it: flat parents with dozens of children, infra/ beside infrastructure/, one new root category per sub-problem, and directories with no summary that no surface can see. kvault guards the write and audits the tree with the same rules, so the two never disagree.

At write time (kvault write --create, MCP kvault_write_node):

• A create that would add a root category is refused unless you pass --new-root.
• A create whose name has the same words as a sibling (release_note beside release_notes) is refused unless you pass --allow-similar. Near-duplicates (crm_prompts beside crm_prompts_legacy, infra beside infrastructure) and a parent pushed past 10 children are reported as a structure note, with the node to read next.
• Missing intermediate parents get a stub summary (a created note; it says "Placeholder" until you rewrite it) instead of becoming an invisible directory.

In the audit (kvault check), one bounded group per prefix, all warn-only:

.kvaultignore at the KB root (one fnmatch pattern per line; a directory pattern covers its subtree) declares the tooling directories and files that are not nodes and are fine. kvault check --code DUPLICATE (repeatable) runs and reports one code; --max-findings 0 and --max-lines 0 return and print its full list. move and delete name the other summaries that still point at the old path (referrer_paths).

One path in. The guards run only on kvault write and the MCP write tools. A node written with a file tool or a shell redirect skips them and never reaches the ops log; on a real KB, 56 of the 59 nodes created in a month arrived that way. Agents and pipelines create nodes through kvault, and check is the backstop for the ones that did not.

Corrections stick. Nobody reviews a KB on a schedule; the owner corrects the agent in use, after the fact. A correction that only lives in a chat is re-proposed the next week, so it is recorded in the node's frontmatter, where every rule reads it: kvault mark <a> --distinct-from <b> (different things; the sibling and duplicate findings and the create guard stop for that pair), kvault mark <parent> --max-children N (this parent is meant to be this wide), kvault mark <parent> --series-ok (this chronology is intentional), kvault mark <node> --verify-by +14d (re-check this node's time-sensitive facts by then). Every plan question carries a default, and a batch may carry --new-root only when it does not increase the root count, so consolidation runs unattended and nothing waits on a person.

The engine: kvault plan turns findings into an ordered worklist with the exact commands — parents over the ceiling are clustered by leading word into new parents, each with a ready kvault move --batch payload; then ghosts, date series, duplicates, sibling collisions, dangling references, loose files, journal drift, stale facts, and summary rewrites. It never applies anything, and the judgment calls (are ml and machine_learning one initiative?) come back as questions. kvault move --batch --confirm runs a JSON list of {from, to} under one lock with one combined propagation list.

$ kvault plan --limit 2
Plan for .: 9 items (showing 2; --limit 0 for all)
1. cluster  projects → projects/atlas
     projects has 46 direct children (ceiling 10); 6 share the leading word 'atlas'
     kvault move --batch --confirm --kb-root /home/me/my_kb <<'EOF'
     [{"from": "projects/atlas_architecture", "to": "projects/atlas/atlas_architecture"}, …]
     EOF
     then: rewrite projects/atlas as a rollup of its 6 children
2. ghost    infra
     no _summary.md — invisible to tree, search, and check
     …
Questions (answer from evidence; defer only when the evidence is not there):
  - projects: are any of these groups one initiative? atlas, billing, ml, crm, support, …

The cadence — every session, nightly, weekly, monthly — lives in skills/kvault-maintenance/SKILL.md, written so a cron or systemd job can load it alone.

kvault check also catches stale propagation, and works as a pre-prompt hook:

{
  "hooks": {
    "UserPromptSubmit": [
      {"type": "command", "command": "kvault check --kb-root /absolute/path/to/my_kb"}
    ]
  }
}

CLI reference

Agent-facing commands accept --json for machine-readable output and --kb-root (auto-detected from cwd by default), before or after the subcommand — as do the output flags -q/--quiet, --explain, --trace, and --strict (see What kvault tells you).

MCP server (optional)

The CLI is the primary interface. For MCP-native clients, a stdio compatibility server ships with the [mcp] extra (Python 3.10+), bound to one KB root per process:

pip install "knowledgevault[mcp]"
kvault-mcp --kb-root /absolute/path/to/my_kb

{
  "mcpServers": {
    "kvault": {
      "command": "kvault-mcp",
      "args": ["--kb-root", "/absolute/path/to/my_kb"]
    }
  }
}

It exposes the same operations as the CLI (kvault_tree, kvault_search, kvault_read_node, kvault_write_node, summary/journal/validation tools, kvault_log_tail for the ops log), plus a strict parent-summary workflow with stale-write detection. Results carry the same did/notes decision reporting as --json, placed before the bulk payload. The write tools (kvault_write_node, kvault_write_entity) accept ancestors="content"|"paths": "paths" (the default since 0.14.0) keeps ancestor_paths but omits the full ancestors[].current_content payload, which can exceed 45,000 characters on a mature KB; pass "content" to inline it. Set KVAULT_ALLOWED_ROOTS to pin allowed roots on shared runtimes. Protocol details: ARCHITECTURE.md.

Every signal above is on the MCP surface too: kvault_check returns the check document (codes=[...] with max_findings=0 for one code's full list), kvault_plan the worklist, kvault_move_entities runs a batch, kvault_mark records a decision (verify_by included), kvault_write_node takes new_root and allow_similar, and kvault_prepare_summary_update returns child gists past the ceiling (children="content" for full bodies). kvault_validate_kb is integrity only.

Reads stay small over MCP (0.16). kvault_search returns compact hits by default (path, title, kind, date, one-line snippet: about 3.5 KB for 10 hits; compact=false for scores and long snippets). parents="gist" on search and reads gives each ancestor's path, title, and first line, about 2 KB for a whole result, where parents="all" used to attach every ancestor's full document to every hit (500+ KB on a mature KB; it is now capped by total_max_chars). kvault_read_nodes reads up to 25 picked hits in one call under one budget that counts whole nodes (8,000 characters by default over MCP; raise it when the client can take more).

It's just files

kvault produces Markdown with YAML frontmatter in a plain directory. No proprietary format, no database to export from. Your existing tools work out of the box:

Python API

from pathlib import Path
from kvault.core import operations as ops

kg_root = Path("my_kb")
outline = ops.build_outline(kg_root, depth=2)          # annotated tree as nested dict
node = ops.read_node(kg_root, "people/contacts/sarah_chen")
result = ops.write_node(kg_root, "people/contacts/new_person", "# Content", create=True)
matches = ops.search_nodes(kg_root, "sarah follow up")

Development

pip install -e ".[dev,mcp]"
pytest -q
ruff check .
black --check kvault/ tests/
mypy kvault/ --ignore-missing-imports

License

MIT