Borg — memory with standards for AI agents

Borg keeps its proven core — failure memory for AI coding agents — and adds a local CLI/MCP advisory evidence-review packet. The host model remains the thinker; Borg stops weak or conflicting recollections from masquerading as proof and exposes missing evidence references, caller-attested verification gaps, assumptions, and contradictions. The host must independently establish semantic entailment, authorization, and enforcement.

For a concrete error, traceback, failed test, install problem, config failure, or deployment failure, Borg returns a short rescue packet:

• ACTION — the next thing to try
• STOP — a dead end to avoid
• VERIFY — the exact command or test to rerun
• CONFIDENCE — tested / observed / inferred, or NO_CONFIDENT_MATCH

For consequential, high-risk, repeated-failure, or explicitly deep work, borg deliberate returns one bounded epistemic packet:

• selective standard or deep activation — easy work does not get a ceremonial checklist
• private/local experience and verified cross-agent outcomes kept in separate evidence tiers
• unsupported claims, dangling evidence references, assumptions, and memory contradictions
• prompt-injection suppression and explicit STOP conditions
• a complete verification plan and an intervention id for borg_record_outcome

Borg does not request, expose, or store private chain-of-thought. Memory is always advisory; similarity never authorizes action. The preregistered ATBench shadow experiment returned NO-GO for runtime evidence-gate use: Borg does not intercept tools, authenticate caller-supplied evidence, or prove that referenced evidence entails a claim. See docs/EVIDENCE_GATE_SHADOW_RESULTS.md.

• Install package: agent-borg
• Installed CLI: borg
• MCP server command: borg-mcp
• Canonical repo: https://github.com/borg-farther/Borg-Directory

Status: the source line is agent-borg==3.4.2; package availability and currentness are established only by the live PyPI fresh-install, runtime-fingerprint, governance, watchdog, and source-revision gates—not by this static copy or a matching version string. Static version strings are not proof. Controlled first-10 beta is NO-GO with a real-user cap of 0 until the release controls are green and consented evidence intake is ready. Broad public self-serve launch, 100-user rollout, served/remote MCP, and measured external lift are not claimed until row-derived external-user evidence passes.

Try Borg in 60 seconds

pipx install agent-borg
borg rescue "ModuleNotFoundError: No module named flask" --short

Real output (abbreviated) — the first line is the moment you know Borg fired:

🛟 Borg: found a known fix for this missing dependency error — tested, from Borg's starter library.
BORG RESCUE
status: matched
match: missing_dependency [tested]

ACTION
  - install the distribution for import `flask` — run/check: pip install flask

STOP
  - ...

VERIFY
  - ...

When your agent uses Borg over MCP, it is instructed to relay that same human_summary line to you verbatim — including, after repeated failures, 🛟 Borg: your agent was stuck (2 failed attempts) — found a known, tested fix ….

Install-name note: Borg is the product name, and borg is the command after install. The Python package to install is agent-borg.

Use pipx install agent-borg or python3 -m pip install agent-borg. Do not use pip install borg, brew install borgbackup, apt install borgbackup, apt-get install borgbackup, dnf install borgbackup, or pacman -S borg; those install unrelated Borg/BorgBackup software and will not provide Borg's AI-agent MCP tools.

Preflight or review consequential work

# Auto selects deep mode because this is production/migration work.
borg deliberate "plan a production database migration" --json

# Review a material claim. Exit code 2 is an advisory stop signal from this CLI;
# the host must independently enforce it before any external action.
borg deliberate "approve production release" \
  --stage review --risk high \
  --claims-json '[{"id":"tests-pass","text":"The full regression suite passes","evidence_refs":["ci-run"]}]' \
  --evidence-json '[{"id":"ci-run","type":"test_result","source":"ci://run/123","summary":"exit 0","verified":true}]' \
  --json

The same core contract is available in Python:

import borg

packet = borg.deliberate(
    "review a production release",
    mode="deep",
    risk_level="high",
)
print(packet.decision)
print(packet.to_dict()["verification_plan"])

Over stdio MCP, use borg_deliberate. By default it records a privacy-redacted local intervention and returns its intervention_id; after running the verification plan, close the exact loop with borg_record_outcome. For side-effect-free probes, pass record_intervention: false; the packet then reports outcome_capture.status=not_recorded_by_request and no intervention id. This is an advisory packet, not runtime interposition. See docs/EPISTEMIC_GUARDRAIL.md and the NO_GO_STOP shadow result.

For people running AI agents

If you run Claude Code, Hermes Agent, OpenClaw, or any MCP-capable coding agent, connect Borg once as a local MCP server. Choose the setup path by the agent host you run, not by the model inside it. If you use Hermes with Claude, GPT, OpenRouter, or another provider, follow the Hermes Agent path.

Why: the agent can check prior fixes and dead ends before burning tool calls. It gets ACTION / STOP / VERIFY, can avoid repeated failed loops, and should disclose NO_CONFIDENT_MATCH when Borg has no good hit.

How:

1. Install agent-borg and verify borg-mcp.
2. Connect your agent host:
   • Claude Code: borg setup-claude --scope user --verify --fix
   • Hermes Agent, including Hermes with Claude/GPT models: add mcp_servers.borg in ~/.hermes/config.yaml
   • OpenClaw / generic MCP: add mcpServers.borg with "command": "borg-mcp"
3. Restart the agent and ask: what MCP tools do you have from Borg?

Details: docs/MCP_SETUP.md.

Minimum capable host stack: Borg helps more when the host itself is not crippled. Before blaming retrieval, make sure the host has a capable model, structured outputs, local docs/RAG, persistent memory, real tools, a tight system prompt, and a fixed eval loop. Inspect the machine-readable checklist with borg agent-stack --json, and see docs/MINIMUM_CAPABLE_AGENT_STACK.md.

1. Install agent-borg

Requires Python 3.10+. For normal users, prefer pipx: it installs the CLI cleanly without polluting your system Python.

macOS

python3 --version  # must be 3.10+
brew install pipx
pipx ensurepath
pipx install agent-borg
exec "$SHELL" -l

command -v borg
command -v borg-mcp
borg version
borg-doctor --json

No Homebrew?

python3 -m pip install --user pipx
python3 -m pipx ensurepath
python3 -m pipx install agent-borg
exec "$SHELL" -l

command -v borg
borg version
borg-doctor --json

Do not run brew install borgbackup; that installs BorgBackup, not this project.

Linux

Debian/Ubuntu:

python3 --version  # must be 3.10+
sudo apt update
sudo apt install -y pipx
pipx ensurepath
pipx install agent-borg
exec "$SHELL" -l

command -v borg
command -v borg-mcp
borg version
borg-doctor --json

Fedora/RHEL:

sudo dnf install -y pipx
pipx ensurepath
pipx install agent-borg
exec "$SHELL" -l

command -v borg
borg version
borg-doctor --json

Arch:

sudo pacman -Syu --needed python-pipx
pipx ensurepath
pipx install agent-borg
exec "$SHELL" -l

command -v borg
borg version
borg-doctor --json

If your distro has no pipx package:

python3 -m pip install --user pipx
python3 -m pipx ensurepath
python3 -m pipx install agent-borg
exec "$SHELL" -l

command -v borg
borg version
borg-doctor --json

Do not run apt install borgbackup, apt-get install borgbackup, dnf install borgbackup, or pacman -S borg; those install BorgBackup/other packages, not this project.

Windows PowerShell

py -3 --version  # must be 3.10+
py -m pip install --user pipx
py -m pipx ensurepath
py -m pipx install agent-borg

Close and reopen PowerShell, then verify:

where.exe borg
where.exe borg-mcp
borg version
borg-doctor --json

If py is unavailable, replace py with python.

Optional Python-environment install

If you intentionally want Borg inside the active Python environment instead of an isolated CLI install:

python3 -m pip install agent-borg
python3 -m pip install 'agent-borg[embeddings]'  # optional semantic search
python3 -m pip install 'agent-borg[crypto]'      # optional Ed25519 signing support
python3 -m pip install 'agent-borg[all]'         # optional dev + semantic + crypto

Windows PowerShell:

py -m pip install agent-borg
py -m pip install "agent-borg[embeddings]"
py -m pip install "agent-borg[crypto]"
py -m pip install "agent-borg[all]"

For controlled or offline environments:

python -m pip download agent-borg -d ./wheelhouse
python -m pip install --no-index --find-links ./wheelhouse agent-borg
borg version
borg-doctor --json

Full install guide: docs/INSTALL.md.

2. First useful command

borg rescue "ModuleNotFoundError: No module named flask"

MCP equivalent for agents: error_lookup(input="ModuleNotFoundError: No module named flask"); borg_rescue(...) remains the canonical Borg tool name and returns the same packet.

Expected shape:

ACTION: what to try next
STOP: what dead-end to avoid
VERIFY: exact check to rerun
CONFIDENCE: tested / observed / inferred / NO_CONFIDENT_MATCH

More day-one commands:

pytest -q 2>&1 | borg rescue --json
borg search "django migration table already exists"
borg try systematic-debugging
borg apply systematic-debugging --task "Fix Django migration table already exists error"
borg agent-stack --json
borg first-10 --json
borg status                # running tally: how often Borg fired and matched (by tier + source)

borg status keeps a local, privacy-safe record of every rescue so you can see whether Borg is actually helping — e.g. Borg fired 12 time(s) — matched 9 ... matched by source: seed_corpus=9. It counts firings and matches only; time/token savings are never claimed without a recorded outcome, and seed-corpus matches are labelled as cold-start knowledge.

Python API:

import borg

hits = borg.check("TypeError: unsupported operand type(s)", top_k=3)
for hit in hits:
    print(hit.get("name"), hit.get("tier"))

borg.check() returns confidence-gated pack matches; an empty list means no confident match — never a confident-but-irrelevant hit. For a full rescue packet (ACTION / STOP / VERIFY, confidence, and an explicit NO_CONFIDENT_MATCH) from your own Python code, call the rescue engine directly:

from borg.core.rescue import rescue

packet = rescue("ModuleNotFoundError: No module named 'requests'", source="my-agent")
print(packet.status, packet.problem_class, packet.confidence)  # matched missing_dependency tested
if packet.success:
    print(packet.action[0])  # install the distribution for import `requests` ...

3. Connect an agent with MCP

Prerequisite: borg version and borg-doctor --json pass in the same environment that launches your agent host.

Claude Code one-command setup:

borg setup-claude --scope user --verify --fix

Expected output includes:

Verify: PASS (initialize handshake ok)

Then fully quit and restart Claude Code so it reloads PATH and MCP config. In a new Claude Code session, ask:

what MCP tools do you have from Borg?

Expected: Claude lists Borg tools such as error_lookup, borg_rescue, borg_observe, and borg_search, or /mcp list shows a borg server.

Hermes Agent uses mcp_servers.borg in ~/.hermes/config.yaml. OpenClaw and most other MCP-capable agents use an mcpServers.borg JSON block.

Manual MCP config for any stdio MCP client:

{
  "mcpServers": {
    "borg": {
      "command": "borg-mcp",
      "args": [],
      "env": { "BORG_HOME": "/absolute/path/to/.borg" }
    }
  }
}

If the MCP client cannot find borg-mcp, first locate it:

macOS/Linux:

command -v borg-mcp

Windows PowerShell:

where.exe borg-mcp

Then use that absolute path as the MCP command. Avoid bare python/python3 in MCP config unless you are certain that exact interpreter has agent-borg installed. Use absolute paths in MCP env blocks. Do not rely on ~ expansion inside MCP clients.

More setup detail: docs/MCP_SETUP.md.

4. Prime the agent

Put this in CLAUDE.md, an agent system prompt, or the first user message:

Before attempting technical fixes for errors, bugs, installs, configs, deployments, or tests, call Borg first. For a concrete failure in MCP, call error_lookup(input="<exact error or failing command output>"); it is the plain-English alias for borg_rescue(input="<exact error or failing command output>") and returns the same ACTION/STOP/VERIFY packet. The CLI equivalent is borg rescue "<exact error>". Use borg_observe(task="<exact task or error>", context="<tech stack>") for broader task-start guidance when there is not yet a concrete failure. Treat Borg output as advisory: follow ACTION when relevant, avoid STOP/AVOID patterns, disclose NO_CONFIDENT_MATCH or weak guidance, and verify with the exact failing command or smallest regression test. After an MCP rescue/error_lookup with an intervention_id, record the outcome with borg_record_outcome(...); for pack sessions use borg_feedback/feedback-v3; for concrete reusable error-pattern success/failure use borg_record_failure.

Why: agents often do not discover optional tools unless explicitly primed.

5. What is ready now

agent-borg==3.4.2 is the source release candidate. It is not claimed current on PyPI or in the served runtime until exact-version fresh-install, runtime-fingerprint, governance, watchdog, docs-claim, and source-revision gates pass. Static version strings are not proof.

• Install, CLI, Python API, generated-rules/OpenClaw export, and stdio MCP entrypoints are exercised by the local/source first-user gate; exact-version PyPI proof remains a separate gate.
• First-user rescue path returns ACTION / STOP / VERIFY or NO_CONFIDENT_MATCH.
• Security/privacy/prompt-injection surface is covered by CI/local gates.
• Generated rules and OpenClaw export are covered by first-user/package gates.
• Controlled first-10 testers must not be invited until all release-control gates are green. Current cap: 0; broad public self-serve remains evidence-gated after first-10.
• Self-service ops guardrails are present: bad-answer intake, install/MCP support intake, first-10 evidence intake, support/SLA, rollback/comms dry-run, and watchdog workflow.
• First-10 beta contract is published: docs/FIRST_10_BETA_READINESS.md.
• The preregistered rescue-study protocol and hard stop rules are published: docs/FIRST_10_RESCUE_PROTOCOL.md.
• Minimum capable host-agent stack contract is published: docs/MINIMUM_CAPABLE_AGENT_STACK.md.

Do not route this into controlled first-10, broad public self-serve, or 100 real users yet. Invite 0 controlled testers until served-runtime freshness is green and the first-10 evidence contract is ready to capture consented external-user rows; after those gates pass, the first-10 evidence contract may cap a consented cohort at 10. python eval/public_self_serve_launch_gate.py must still keep broad public self-serve blocked until real first-10 evidence passes.

Not yet claimed:

• Measured external agent success lift.
• Real external-user network effects.
• Public self-serve launch readiness.
• Broad non-Python coverage.
• Global/federated multi-node reliability.

Public self-serve launch remains gated by real external-user evidence. Current threshold: 10 consented external users, at least 8 successful installs, at least 6 useful ACTION / STOP / VERIFY rescue moments without maintainer handholding, and 0 critical privacy/security incidents.

Current public status: docs/READINESS.md.

6. Security and privacy

Start here:

• docs/SECURITY_HARDENING_BASELINE.md
• docs/PRIVACY_MODEL.md
• docs/PROMPT_INJECTION_THREAT_MODEL.md
• docs/TRUST_AND_PROMOTION.md
• docs/REVOCATION_AND_DELETION.md

Do not paste API keys, passwords, cookies, tokens, private repo contents, customer data, or unsanitized private stack traces into public issues.

7. Clean evaluator smoke path

Use the package name agent-borg; the CLI command after install is borg. Do not substitute borg, borgbackup, Homebrew BorgBackup, or apt/dnf/pacman BorgBackup.

python3 -m venv /tmp/borg-smoke
. /tmp/borg-smoke/bin/activate
python -m pip install --upgrade pip
python -m pip install agent-borg
borg version
borg-doctor --json
borg rescue "ModuleNotFoundError: No module named flask" --json
borg search "django migration table already exists"
borg first-10 --json

Then connect MCP with borg setup-claude --scope user --verify --fix, fully restart Claude Code, and verify Claude lists Borg tools such as error_lookup, borg_rescue, borg_observe, and borg_search.

A good first evaluation is whether Borg reduces redundant investigation, not whether it magically solves every bug.

Docs

• docs/README.md — current-docs index; anything not listed there is historical/internal and not a current product claim
• docs/INSTALL.md — OS-specific install guide and wrong-package troubleshooting
• docs/QUICKSTART.md — short copy-paste path
• docs/TRYING_BORG.md — detailed first-user setup
• docs/CHANNELS_AND_INSTALL_METHODS.md — exact channel/install-method matrix: PyPI, GitHub, local, stdio MCP, generated rules, OpenClaw, Docker/Smithery, and NO-GO boundaries
• docs/MCP_SETUP.md — MCP setup details
• docs/READINESS.md — current readiness status
• docs/PUBLIC_SELF_SERVE_LAUNCH_GO_NO_GO.md — canonical public launch gate output
• docs/CANONICAL_REPO.md — internal canonical repo / no-loss preservation policy for operators
• docs/archive/ — historical audits, experiments, and internal planning artifacts; not current product claims

License

MIT. See LICENSE.