Agent memory that thinks in
tokens, not vectors

Embedding-based memory has
three structural flaws —
and three costly consequences

Every major agent memory system in 2026 converts conversations to dense vectors before retrieval. This works for factual lookups — and breaks for the queries agents most frequently need to answer.

Stay in token space,
end to end

The most valuable signals in conversational memory are structural, not semantic: what kind of memory did the user express? and does this episode's memory type match what this query needs? These questions are answerable with token-level pattern matching — no embedding model required.

Lower cost per
correct answer

RAG cost is not just search cost. It is every oversized prompt, every rerank, and every retry needed before the model finally sees the right evidence. ContextFit is designed to retrieve less, better — reducing the context loops between question and correct answer.

Public memory retrieval,
measured like gbrain.

LongMemEval-S session retrieval asks one clean question: did the system put a ground-truth answer session in the top 5? No answer generation, no LLM judge, no hidden scoring model. Answer-quality benchmarks stay separate because they depend on the downstream model and judge.

Strengths where AI agents most need them. The hardest memory failures are rarely simple keyword lookups — they are open loops, preferences, temporal changes, and cross-session synthesis. These are the moments where better retrieval reduces retries, preserves user trust, and turns memory into useful action.

No database. No GPU.
No vendor lock-in.

The token-native architecture isn't just a performance choice — it's a deployment choice. ContextFit runs anywhere a Python process runs.

Chunks follow document meaning,
not blind token windows

ContextFit still stores and searches token IDs, but file ingestion now chooses smarter boundaries first: Markdown headings, prose paragraphs, TMD ledger rows, JSON events, CSV/TSV records, email messages, calendar events, and source code symbols. The result is more auditable evidence with richer per-chunk metadata.

A tiny ledger format agents can actually trust

TMD ledger is a new ContextFit-proposed Tabular Markdown file format: a human-readable text file for records that should stay row-addressable, schema-aware, and easy to cite. It bridges Markdown notes and structured data without requiring a database.

Simple, composable,
fully interpretable

Drop in as a memory layer. Query with a natural-language string. Get back ranked session IDs with source-linked evidence.

# Ingest sessions — no API calls, no GPU, ~4ms per session
from contextfit import RetrievalEngine

engine = RetrievalEngine()
engine.ingest_sessions(sessions)
engine.save("./memory_index")

# Query — auto-routes to the right retrieval mode
result = engine.query_auto(
    "what should I cook for dinner tonight?",
    top_k=5
)

# Returns ranked session IDs + route metadata
print(result["route"])        # → episode_score
print(result["session_ids"])  # → ["s_garden_harvest", ...]

# Or use individual modes directly
engine.rank_sessions_by_episode_score(query, top_k=10)
engine.query(query, method="hybrid", top_k=50)
engine.rerank_sessions_by_structure(query, bm25_order, session_texts)