Skip to main content
Glama

foghorn

Decision staleness alerts for AI agents.

foghorn

CI PyPI version Python 3.10+ Downloads License: MIT codecov Typed

Quick Start · How It Works · CLI Reference · GitHub Action · vs. Alternatives · Contributing


Why

AI agents make decisions. Those decisions depend on facts about the world. The world changes.

When the facts an agent depended on are no longer true, its conclusions become stale — but nothing tells you which ones, or how much to worry. You either re-run everything (expensive) or trust outdated conclusions (dangerous).

foghorn solves this by treating agent knowledge like source code: every fact and decision is version-controlled, content-addressed, and diff-able. When facts change, foghorn tells you exactly which decisions are affected and how confident you should be about the impact.

foghorn stale --exit-code   # Fails CI if any agent decision is based on stale facts

Related MCP server: adr-skills

How It Works

flowchart LR
    A[Agent records Fact\nsubject · predicate · object] --> B[Agent records Decision\ndepends_on Fact IDs]
    B --> C[foghorn commit\nWorldCommit snapshot]
    C --> D{Fact changes\nnew triple added}
    D --> E[diff_commits\ndetects added/removed]
    E --> F[compute_staleness\nfinds affected decisions]
    F --> G[StalenessAlert\nimpact_score ranked]

Core primitives:

  • Fact — an immutable, content-addressed triple (subject, predicate, object). ID = SHA-256[:16] of the triple. Two agents recording the same fact always get the same ID.

  • Decision — a named agent conclusion that records which Fact IDs it depended on.

  • WorldCommit — a snapshot of all facts and decisions at a point in time.

  • StalenessAlert — emitted when a decision's upstream facts have changed, ranked by impact_score (confidence-weighted).

Facts and decisions are staged, then committed in batches — exactly like git. diff_commits() computes the fact-level delta between two commits, and compute_staleness() propagates that delta through the dependency graph in O(changed_facts × avg_decisions_per_fact).


Features

Feature

Details

Content-addressed facts

Same triple always produces the same ID — no duplicates

Decision dependency graph

Decisions explicitly declare which facts they relied on

Staleness propagation

compute_staleness() finds all affected decisions in one pass

Confidence-weighted impact

impact_score reflects how certain the now-stale facts were

Offline / local-first

Single SQLite file, no server required

CI exit code

--exit-code makes foghorn stale fail CI if anything is stale

JSON output

Machine-readable output for downstream automation

Markdown output

Ready-to-paste GitHub PR comment

FastAPI REST server

/fact, /decide, /commit, /stale, /log endpoints

MCP server

Model Context Protocol integration for Claude and other agents

114 tests

Comprehensive test suite covering all layers


Quick Start

pip install foghorn-ai
from foghorn.repo import WorldRepo

repo = WorldRepo.init(".foghorn/world.db")

# Record facts your agent is relying on
f = repo.add_fact("Redis", "is-appropriate-for", "rate-limiting", confidence=0.95)
pg = repo.add_fact("Postgres", "is-primary-db", "yes")

# Record a decision that depends on those facts
repo.decide(
    "chose-redis-for-rate-limiting",
    "Redis is fast enough for our rate-limiting needs at current scale.",
    depends_on=[f.id],
)

commit = repo.commit("Initial architecture decisions")
print(commit.id)  # e.g. "a3f8b2c1d4e5f6a7"

# Later — the world changed: retract the old fact, add the replacement
repo.retract_fact(f.id)
repo.add_fact("Redis", "replaced-by", "Valkey")
repo.commit("Redis EOL notice")

# Which decisions are now stale?
alerts = repo.stale()
for alert in alerts:
    print(f"STALE: {alert.decision_label} (impact: {alert.impact_score:.0%})")

CLI Reference

foghorn [--db PATH] COMMAND [OPTIONS]

Command

Description

Key options

fact SUBJECT PREDICATE OBJECT

Stage a new fact triple

--confidence FLOAT

decide LABEL CONTENT

Stage a decision

--on FACT_ID (repeatable)

commit

Commit all staged items

-m MESSAGE (required)

stale

Show stale decisions

--since COMMIT_ID, --format {rich,json,markdown}, --exit-code

diff

Show fact changes between HEAD and parent

--format {rich,json,markdown}

log

Show commit history

status

Show staged item count and HEAD

recommend

Show actionable recommendations for all stale decisions

Global options:

Option

Default

Env var

--db PATH

.foghorn/world.db

FOGHORN_DB

Examples:

# Stage facts
foghorn fact Redis is-appropriate-for rate-limiting
foghorn fact Postgres is-primary-db yes --confidence 0.9

# Stage a decision that depends on the Redis fact
foghorn decide chose-redis "Redis fits our rate-limiter requirements" \
    --on a3f8b2c1d4e5f6a7

# Commit
foghorn commit -m "Initial architecture decisions"

# Check for staleness (machine-readable)
foghorn stale --format json

# Fail CI if anything is stale
foghorn stale --exit-code

GitHub Action

Add foghorn staleness checks to your CI pipeline:

# .github/workflows/foghorn.yml
name: foghorn staleness check
on: [push, pull_request]

jobs:
  stale:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: sandeep-alluru/foghorn@main
        with:
          db: .foghorn/world.db
          fail-on-stale: "true"

The action installs foghorn, runs foghorn stale --exit-code, and fails the job if any decisions are stale. See docs/github-action.md for full documentation.


vs. Alternatives

foghorn

Graphiti / Zep

Letta / Mem0

Memoria

LangGraph checkpointing

Decision-dependency tracking

Yes — explicit fact IDs per decision

No

No

No

No

Staleness alerts

Yes — ranked by impact_score

No

No

No

No

Content-addressed facts

Yes — SHA-256[:16]

No

No

No

No

Offline / local

Yes — single SQLite file

Requires Neo4j/Redis

Requires server

No

Partial

CI exit code

Yes — --exit-code flag

No

No

No

No

Primary purpose

Decision staleness tracking

Long-term agent memory

Personalized memory

In-context memory

State persistence

Graph storage

Dependency edges only

Full knowledge graph

Vector + metadata

In-context only

State snapshots

Open source

MIT

Open core

Open core

MIT

Apache 2.0

foghorn is not a general-purpose agent memory system. It is specifically designed to answer: "Given that these facts changed, which agent decisions are now invalid?"


Claude / MCP integration

foghorn ships a Model Context Protocol server that lets Claude and other MCP-compatible agents record facts and decisions directly:

# Start the MCP server
python -m foghorn.mcp_server

# In your Claude Code project's .claude/settings.json:
{
  "mcpServers": {
    "foghorn": {
      "command": "python",
      "args": ["-m", "foghorn.mcp_server"]
    }
  }
}

Once connected, Claude can call foghorn/fact, foghorn/decide, foghorn/commit, and foghorn/stale as tools. See docs/mcp.md for the full tool schema.


OpenAI integration

foghorn exposes a FastAPI REST server compatible with OpenAI's function-calling format. The tool definitions are in tools/openai-tools.json and the full API spec is in openapi.yaml.

# Start the REST server
uvicorn foghorn.api:app --reload

# Pass to Codex CLI or any OpenAI-compatible agent
codex --tools tools/openai-tools.json "Check which architecture decisions are stale"

Endpoints: GET /health, POST /fact, POST /decide, POST /commit, GET /stale, GET /log. See docs/openai.md for details.


Case Studies

See how teams are using foghorn in production:


Repository structure

foghorn/
├── src/
│   └── foghorn/
│       ├── fact.py           # Fact, Decision, StalenessAlert dataclasses
│       ├── store.py          # SQLite-backed WorldStore + WorldCommit
│       ├── staleness.py      # DiffResult, diff_commits(), compute_staleness()
│       ├── repo.py           # WorldRepo high-level API
│       ├── report.py         # print_stale(), print_diff(), to_json(), to_markdown()
│       ├── export.py         # export_json(), import_json(), export_graphviz()
│       ├── propagate.py      # propagate_staleness(), PropagationResult
│       ├── recommend.py      # recommend(), Recommendation
│       ├── cli.py            # Click CLI (fact, decide, commit, stale, diff, log, status, recommend)
│       ├── api.py            # FastAPI REST server
│       └── mcp_server.py     # MCP server (list_facts, record_decision, commit, check_stale)
├── tests/
│   ├── test_fact.py          # Fact, Decision, StalenessAlert unit tests
│   ├── test_store.py         # WorldStore + WorldCommit tests
│   ├── test_staleness.py     # Staleness propagation tests
│   ├── test_repo.py          # WorldRepo integration tests
│   └── test_cli.py           # CLI subprocess integration tests
├── examples/
│   └── demo.py               # Standalone demo script
├── docs/                     # MkDocs documentation
├── tools/
│   └── openai-tools.json     # OpenAI function-calling tool definitions
├── assets/
│   ├── hero.png              # README hero image
│   └── logo.png              # Project logo
├── action.yml                # GitHub Action
├── openapi.yaml              # OpenAPI 3.1 spec
├── pyproject.toml            # Package metadata + dependencies
└── CONTRIBUTING.md           # Contribution guide

GitHub Topics

Suggested topics for discoverability:

ai-agents decision-tracking staleness-detection knowledge-graph sqlite mcp openai langchain llm-tools agent-memory fact-tracking ci-cd python


Star History Chart


Stay Updated

Subscribe to The Silence Layer — weekly dispatches on production AI infrastructure, new releases, and the failure modes that production AI systems don't surface until it's too late.

A
license - permissive license
-
quality - not tested
B
maintenance

Maintenance

Maintainers
Response time
Release cycle
Releases (12mo)
Commit activity

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/sandeep-alluru/foghorn'

If you have feedback or need assistance with the MCP directory API, please join our Discord server