wikimemFile-first memory for AI agents

Categories + wiki-links over plain markdown. No database, no embedding model, no docker — pip install wikimem and it works.

Getting Started

What is wikimem?

GitHub

📄

Markdown is the database

One file per category, one

🔗

Wiki-links recall what search misses

[[category:item]] links written at memorization time connect meaning-related items that share no words. Expansion is a mechanical one-hop lookup — no LLM, no graph database.

🔍

Zero-dependency BM25

In-memory index, rebuilt free at startup. Chinese works out of the box via character bigrams; install [zh] for jieba.

🧭

Semantic fusion, strictly optional

The [embed] extra fuses BM25 with cosine similarity. BM25 is never disabled — if the endpoint is down, retrieval silently carries on.

⚡

Never blocks the conversation

Retrieval is synchronous, token-budgeted, and makes zero LLM calls. Memorization is async and costs the host at most one.

🧾

Always answerable

Every mutation appends one line to journal.jsonl; every retrieval can explain its scoring. tail -f is your observability stack.

Sixty-second taste

python

from wikimem import MemoryIndex, MemoryStore

store = MemoryStore("memory/")
store.add("preferences", "likes-the-sea",
          "Loves the seaside, mentioned wanting a beach trip. [[daily_life:beach-trip-plan]]",
          owner="user:xnne", source_conv="conv_20260710")
store.add("daily_life", "beach-trip-plan", "Planning a summer beach trip to watch the sunrise.")

index = MemoryIndex(store)  # in-memory BM25, rebuilds itself on store writes
result = index.retrieve("beach vacation", budget_tokens=800)
for entry in result.items:
    # hits come ranked; each is followed by its one-hop wiki-link targets
    print(entry.source, entry.item.name, entry.score)

What's on disk afterwards? Two markdown files you can open in any editor, and a one-line-per-mutation journal. That's the whole system.