Head to head

ArgosBrain vs Aider

Aider names symbols. ArgosBrain understands them.

What Aider does
Stateless PageRank over tree-sitter surface names, packed into a ~1k-token repo map per request.
What ArgosBrain does differently
Persistent, compiler-grade semantic graph with members, overrides, and cross-file resolution.
What Aider is

The baseline, stated fairly.

Aider (Apache-2.0) has no persistent memory. From their own docs: "when you quit and restart aider, it has no memory of what you discussed, what architecture decisions you made, or what conventions you established."

Its Repository Map is a stateless, live-computed summary: tree-sitter extracts top-level symbol names across 130+ languages, files-as-nodes/references-as-edges form a graph, PageRank ranks importance, and a binary search packs top-ranked symbols into a token budget (default 1000 tokens) sent with every request.

How it actually works

Technical facts.

Sources: Repo map docs · Conventions docs

Depth vs breadth — important

130+ languages ≠ semantic understanding.

That number describes tree-sitter grammars available for surface-level name extraction — not semantic understanding. Tree-sitter produces a concrete syntax tree for a single file; it does not perform cross-file symbol resolution, type inference, import resolution, or method-override analysis. Aider's repo map is therefore closer to "a ranked list of file names and the top-level identifiers in them" than to a true code graph.

ArgosBrain uses a tiered ingestion pipeline that picks the highest-precision technology available per language: compiler-grade SCIP where a mature indexer exists, live LSP next, bespoke tree-sitter drivers with semantic hooks where neither exists.

See the FAQ on language-server failure modes (Q1–Q3).

QueryAider repo mapArgosBrain
resolve_member("foo.bar.baz")✗ No member resolution✓ Exact
Every method on class X✗ Not modeledcontained_by traversal
Every caller of X.method✗ Approximate (text-match)✓ Call-graph traversal
Every override of Interface.method✗ No inheritance graph✓ Inheritance + override edges
Who reads DATABASE_URL?✗ grep-equivalent✓ Config-surface traversal
symbol_exists("teleport_to_mars")✗ Can't distinguish "absent from map" from "doesn't exist"✓ Deterministic no
Verdict

Where each one wins.

↑ Where ArgosBrain wins
  • Depth over width. Aider's 130-language figure covers surface name extraction. The head-of-distribution languages where 95% of code is written are exactly where ArgosBrain runs compiler-grade SCIP / live LSP / semantic-hook tree-sitter.
  • Persistence. Aider forgets everything at quit.
  • $0 retrieval. Aider's 1k-token repo map ships on every request.
  • Cross-session learning — verify / dispute / strengthen. Aider has no notion.
↑ Where Aider wins
  • Tree-sitter PageRank is elegant for stateless context packing; Aider is the inspiration for part of our ingest.
  • Long-tail coverage. If you're on an esoteric language at the far end of the distribution, Aider's tree-sitter grammar list is broader than ours. Trade-off: you get a file + name list, not a semantic graph.
  • Proven CLI UX — one of the most popular OSS coding agents.
When to choose which

Honest recommendation.

Choose Aider if
  • You want a CLI coding agent with no setup
  • You don't need memory
Choose ArgosBrain if
  • Aider + ArgosBrain (via MCP): Aider for the agent, ArgosBrain for the memory Aider explicitly doesn't provide