AI coding agents are amnesiac and ungoverned. Every task restarts by re-reading the same files, long sessions drift onto stale assumptions, and nothing warns the agent when a change is about to break a contract or cross a boundary.
OpenLore fixes both halves, and then gets out of your way. A one-time static analysis builds a live knowledge graph of your repo — call structure, types, tests, decisions, IaC. From then on your agent is already oriented at the start of every task and its changes are — automatically, deterministically, with , so the same question returns the same grounded answer and the agent is instead of served a confident guess.
Persistent architectural memory and structural cognition for AI coding agents.
2026-11-14
ดาว
—
เครื่องมือ
—
อัปไทม์
—
certified before they land
no LLM in the hot path
told when a fact goes stale
Install once, then forget it
npm install -g openlore && openlore install
That is the only command most people ever run. It auto-detects your agent (Claude Code, Cursor, Cline, Continue, Pi, AGENTS.md), wires it, registers the MCP server, and builds the index — no API key, no config, no prompts. After that, it's on in the background and you don't think about it:
🔄 Every task starts oriented, automatically — a compact architecture digest is injected into your agent's context at the start of each session, and your agent calls orient() and the other tools itself over MCP. You don't type anything.
🌱 Every repo, not just this one — install wires your user account too, so each new git repo you open indexes itself in the background on first touch. "Install once" is literal.
🩹 Always fresh — a file watcher re-indexes as you edit and converges to a full analysis; anything it can't recompute in time is marked explicitly stale, never silently wrong.
🛡️ Changes get checked before they land — the governance tools (and openlore review on a PR) run deterministically with no API key.
The CLI below exists for humans and CI — prove, review, doctor, features. The everyday path is: install, then let your agent use it.
Zero config. Run openlore features to see every opt-in capability (embeddings, the commit gate, the spec store…), whether it's on, and the one command to enable it. openlore doctor verifies the wiring.
What you get
Two things, both deterministic and local:
🧠 Memory — your agent starts every task already oriented
Orient in one call — orient(task) returns the relevant functions, callers, specs, tests, and insertion points (~430µs p50 on a 15k-node graph) instead of a dozen exploratory file reads.
Survives refactors — anchored notes and decisions carry forward when a symbol is renamed or moved, instead of orphaning.
One graph for everything — application code, Infrastructure-as-Code, and architectural decisions share one graph, so a single query spans all three.
Told when a fact is stale — every response carries a brief, factual freshness note (the Epistemic Lease) once a cached fact has aged or the repo moved, so a long session can't drift onto a confident guess.
🛡️ Governance — guardrails on what the agent changes
Breaking-change verdict — certify_public_surface classifies every changed export breaking / non-breaking / potentially-breaking and names the consumers each break hits. Never silently "safe."
Sensitive-boundary check — change_impact_certificate flags when a diff opens a new path into a boundary you declared (reachable after the change, not before).
Grounded claims — verify_claim returns confirmed / refuted / unverifiable with a citation, before an agent asserts "X is dead" or "Y is safe to change."
One commit gate — openlore enforce blocks on findings you mark blocking. Advisory by default, no API key.
The one line to remember: an agent's expensive failure mode isn't ignorance — it's confidence. A model that doesn't know a function exists will go look. A model that "knows" a stale fact builds on it, and you pay at review time. OpenLore is built so the agent can be told "that fact is stale" and "this change opens a path you said was sensitive" — deterministically, with no second model guessing about the first.
Value Scorecard — does it pay for itself?
OpenLore only earns its place if an agent with it reaches a correct answer for less total cost than the same agent without it. We measure that and publish it — wins and losses. Numbers from the Spec 14 benchmark (claude -p, sonnet, N=4 medians, pinned SHAs), measured 2026-06-01.
Scenario
Cost Δ
Round-trips Δ
Verdict
Large/unfamiliar repo · deep "how does X flow through Y"(its target)
−7% to −21%
−26%
✅ helps — and the win grows with repo size
Small/familiar repo · shallow "who calls X"
task-dependent (Round 1: +43%)
+38%
❌ often adds overhead — measure first
The win scales with codebase size (round-trips without → with):
Repo (size)
Cost Δ
Round-trips
excalidraw (~640 files)
−21%
25 → 16
tokio (~790 files)
−21%
17 → 13
okhttp
−13%
13 → 11
django (~3k files)
−7%
21 → 15
gin (110 files, smallest)
+4% (≈even)
10 → 9
Re-confirmed live 2026-06-03: the deep-task win reproduces (okhttp −13%); the small/familiar case is task-dependent, not a flat loss (chalk −32% win vs. express +59% loss). Don't guess from our repos — run openlore prove --estimate on yours (seconds, no API key).
Honesty contract. We never publish a savings number the benchmark didn't produce, we always show the losses next to the wins, and every claim traces to a command you can run. Methodology: docs/AGENT-BENCHMARKS.md.
Is OpenLore for you?
✅ Strong fit
A codebase too big to hold in your head — and the model's. Private or niche code the model never memorized. Long sessions where stale assumptions compound. Polyglot repos, or code plus the IaC that deploys it.
🤔 Try it, measure
Mid-size repos and mixed workloads. Run openlore prove --estimate (seconds, no key) before you commit.
❌ Probably not
A small repo the model already knows, answering shallow questions — your agent's built-in search is cheaper, and we publish the measurement that says so. Also: OpenLore locates and certifies, it never edits your code.
What your agent reaches for (you rarely type these)
Once wired, your agent calls these over MCP on its own. The default surface is the substrate preset — 15 tools: the navigation core, the highest-value governance reads (recall, verify_claim, blast_radius), and the spec composites. The lean navigation preset (10 tools) and the full 80 tools (--preset full) are one flag away.
When the agent needs to…
Tool
Start any task
orient(task) — functions, callers, specs, insertion points in one call
Everyone in this category answers the same first question: how does the agent see the codebase without reading it file by file? LSP toolkits answer with symbols, graph MCP servers with a parsed graph, search platforms with an index. OpenLore answers it too — then keeps going into the second question almost nobody answers: what happens when the agent starts writing? A retrieval layer makes an agent informed; it doesn't make it safe. That second half — governance on the same graph, no LLM in the loop — is what OpenLore is for.
Agent built-ins
LSP toolkits
Graph MCP
Search platforms
OpenLore
Structural context instead of file reads
❌ grep
✓ symbols
✓ graph
✓ index
✓ call graph + IaC + decisions on one graph
Local, no API key, deterministic
Partial
✓
✓
❌ hosted
✓ no LLM in the hot path
Cross-session memory anchored to code
❌
Partial
✓ notes
❌
✓ carried across renames
Told when a cached fact goes stale
❌
❌
❌
❌
✓ Epistemic Lease
Breaking-change verdict over a diff
❌
❌
Partial
❌
✓ per export, consumers named
"Did this diff open a sensitive path?"
❌
❌
❌
❌
✓ differential, pre-commit
Cost effect published with the losses
❌
❌
❌
❌
✓ −26% round-trips on deep tasks
Use the other tool when it's the right one: symbol-level edits (rename across files) are an LSP toolkit's home turf — OpenLore is deliberately read-only, and the two compose. Org-wide search across hundreds of repos is a search platform. Just-fast-graph-retrieval is a smaller graph MCP server. And a small, familiar repo with shallow questions is cheaper with your agent's built-in search. OpenLore also exports SCIP, so it sits alongside these tools, not against them.
Comparisons reflect each project's publicly documented capabilities as of July 2026; a correction PR is always welcome.
27 MB under .openlore/ (gitignorable; rebuildable from source)
Per-query latency
~430 µs p50 in-process via the MCP server
Your source code
never leaves the machine — no account, no telemetry (opt-in only)
Lock-in
none — delete .openlore/ and nothing about your repo has changed
Large monorepos take minutes rather than seconds, stated plainly in Known limitations. Platform: Linux and Windows are exercised in CI; macOS is supported.
Fully local — nothing phones home. No account, no network call; your source, the index, and every query stay on your machine. The only diagnostics that exist are opt-in and local-only, off by default and never transmitted: OPENLORE_TELEMETRY=1 records tool calls, agent identity, latency, error messages, decision titles, and lease events to the gitignored .openlore/telemetry/ directory, and OPENLORE_LLM_LOGS=1 stores prompts and responses (after secret redaction) in the gitignored .openlore/logs/ directory. → docs/configuration.md
Languages & Infrastructure-as-Code
Languages: TypeScript · JavaScript · Python · Go · Rust · Ruby · Java · C++ · Swift · C# · Kotlin · PHP · C · Scala · Dart · Lua · Elixir · Bash — call graphs ride the same primitives for every language. → docs/language-support.md
Infrastructure-as-Code: Terraform/HCL · Kubernetes · Helm · CloudFormation · Ansible · Pulumi · AWS CDK · CDKTF · Dockerfile · Docker Compose · GitHub Actions · Azure Bicep — IaC resources project onto the same graph as application code, so orient, search_code, and analyze_impact answer "what's the blast radius of changing this security group / IAM role / base image / CI job?" with zero new tooling. → docs/iac.md
Known limitations
We'd rather you know these up front.
Static analysis only. Polymorphic dispatch, event channels, and route→handler bindings are recovered (provenance-labeled synthesized). What isn't: reflective invocation with a non-literal target (getattr(o, name)()), eval, and DI/plugin registries with no visible binding — silently absent from the graph.
LLM spec quality varies. Generated spec prose reflects the model's understanding — review complex business logic. Structure, format, coverage, and drift are checked deterministically.
Keyword (BM25) search is the default; semantic is an opt-in upgrade.orient/search_code work immediately with no key. Add hybrid dense+BM25 with openlore embed --local (bundled, CPU-only, ~23 MB).
Large monorepos analyze one package at a time after the first full build (openlore analyze --shard <name>); the initial index is still a full analysis.
Incremental updates converge or flag, never silently diverge. When a change exceeds the per-save budget, the rest is marked explicitly stale and self-heals on later edits or a background re-analyze.
Documentation
Start here: the documentation index maps every task to the one canonical page that answers it. The essentials:
If OpenLore saves your agents from re-reading the same files — or catches one risky change before it lands — star the repo. It's the signal that tells us to keep building, and it helps other engineers find it.
