andenken

Let past conversations gain meaning in the present. Files are the authority. The timeline is the compass. A semantic memory axis for recovering the reasons behind a decision — and where to pick up next.

Find it so you can continue

Why did we choose that design? Where did I write about this? What followed that decision?

As conversations grow and harnesses change, a surviving record can become hard to reach. andenken does not keep an old session running. It helps you choose which original record to open: find candidates through today’s question, check their sources, read what matters, and take the next step.

Andenken means recollective thinking — thinking toward what has been. We call it 뜻새김 in Korean: letting the past gain meaning again, rather than pouring every old conversation back into a prompt.

When to come here

• Continuing work in a different session: recover not just the conclusion, but the reasons and conditions behind it, then return to the original conversation.
• Connecting thoughts and writing: find public notes written in different words, and neighbours of the idea you are exploring now.
• Comparing human and bot records: keep track of who left evidence on which surface, rather than mistaking one source for the whole memory.

Reconstructing an entire day is a job shared with the timeline and other tools. andenken does not collect every record into an all-in-one personal brain.

Three retrieval surfaces, named sources

The engine combines vector retrieval with BM25; available retrieval legs depend on the calling surface and index health. OpenClaw search currently uses the CLI’s search-openclaw path; there is no dedicated Pi tool. Its harvest reuses compatible vectors, so the import itself makes zero re-embedding API calls — not a claim that upstream embedding or queries are free. Provenance and ownership remain distinct.

A small retrieval engine, with authority outside it

Admitted conversations / public writing → chunk → embed → retrieval stores
Already computed bot text and vectors ─────────────────→ a separate store

Question → choose surface and filters → vector + BM25 → fusion / diversity
         → candidates with sources → read selected originals

Conversations and garden notes currently use Qwen3-Embedding-8B / 4096d, stored in LanceDB. Vectors find related expressions; BM25 finds concrete words. Candidates are fused and re-ranked for diversity. Recent-first conversation lookup has its own stored-signal path; ordinary relevance does not decay a memory out of reach merely because it is old.

Meaning crosses languages. GLG writes in Korean, while the question that looks for those records may arrive as an English term or an agent’s English summary. Word matching alone is not enough to cross that wall, so conversations and notes live in a multilingual embedding space where the same meaning, written in another language, can still land nearby. This is the direction of the design, not a promise of cross-lingual quality; personal vocabulary and morphology stay with their own tools.

Records outlive the index. The session corpus merges admitted files across devices, append-only, with a checksum manifest and roster preserving provenance. The sessions query replica receives index, manifests and corpus together. Neither one machine’s storage nor one model’s context window should become the owner of memory.

Small enough to keep. An index that grows toward ten gigabytes per checkout cannot be carried for a lifetime, so size now comes before retrieval polish. The sessions index is built under a pinned selection policy: lifecycle notices injected by the harness and sibling-mail doorbell envelopes are dropped before embedding, and messages containing credential-shaped strings are held rather than sent. Only the searchable trace is selected; no original conversation is deleted. A freshly built store came out far smaller; the dropped rows explain only a small part of that, and most of the reduction appears to come from rewriting accumulated storage — the change history keeps the measured scope.

From time to meaning, and back

The timeline is the skeleton; the embedding is the lens.

1. Time → meaning: start with a date or event, then recover conversational decisions and garden interpretations around its canonical coordinates.
2. Meaning → time: start with a question, find related records, then return their dates, files and identifiers to the timeline to check the surrounding context.

The harness-side timeline owns when and what happened. An embedding similarity score is neither evidence of a date nor a probability of truth.

Search offers rooms; reading opens one

Question → start with a few candidates → check sources → open selected originals → decide

The first screen helps you choose a record instead of flooding context with chunks. Narrow conversations by source, project and time window; return to public notes through titles and Denote IDs. Use structural lookup to establish that a particular note exists, and the timeline to establish when an event happened. With source=all, a sparse conversation search may append a few public-note results, labelled as supplementary sources.

When a result looks wrong, or two traces disagree, return to the original conversation before tuning the index. An excerpt command reads the lines around a hit directly from the source file. A trace points the way to the original; it never stands in for it.

The same distinction matters for pi-durable: the embedded selection is not the whole conversation. Reading a closed snapshot has its own path; embedding follows an explicit admission policy and checkpoints. andenken neither silently opens a live conversation database nor claims to remember every latest turn.

Who owns the memory?

Session originals live in a separate, private git-annex archive: agent-config, the harness configuration repository, owns its storage method, version retrieval and ledger/commit rules, while andenken implements its own gathering, admission, manifests, replication and reader, and builds the derived index. Implementing a reader is neither unlimited permission to read nor a certificate that what it reads is true.

Global publication checkpoints that archive before indexing and checks its source receipt before the index ships. Corpus, manifests and index stay coherent; changed inputs are handled incrementally. This project does not run automatic dreaming or overnight consolidation by default.

When in doubt, go back to the originals. Because the session originals are preserved, the sessions search index can be rebuilt from them; if that store ever drifts into something strange, re-embedding from the originals is a real way out. That is a strategic option, not a standing permission: a full re-embedding is paid work and, like global publication, still needs its own human decision.

Depending on the configured provider, embedding uses an external API. Public source code does not make personal conversation corpora public. Operators must decide what to collect, what to send to a provider, and what spending to authorize.

Zero results and retrieval failure are not the same observation. Index integrity, retrieval behaviour and human usefulness are separate acceptance layers. The README and change history record the implementation and measured boundaries.

The direction we keep

• Keep the trace small enough to carry: prefer a store that can last for years over improvements that cannot be sustained. Improvements are tried within that size; if the sessions store starts to look wrong, the way out is a fresh embedding from the preserved originals, decided by a human.
• Keep conversation admission explicit: the sessions index follows a pinned selection policy. Changing what counts as a meaningful human–agent exchange stays a deliberate decision, backed by dry runs and evidence.
• Let quality follow real work: perfect retrieval is not today’s gate. When quality returns to the front, judge it by timeline-grounded tasks — evidence rank, provenance, diversity and freshness gaps — rather than keyword hits, and ask whether retrieval helped the next decision.
• Reduce calling-surface gaps: expose the differences between CLI and Pi reach, and treat tool access for surfaces such as OpenClaw as separate follow-up work. Leave pre-reply recall orchestration, session transport, time coordinates, structural navigation and vocabulary to their owners.
• Explore source structure: the disabled Org track remains upstream research. Reactivation or further vocabulary/structure integration requires its own demonstrated need and verification.

These are directions, not completed features or release promises. Repository documents keep concrete next steps and open questions.

Beside the house that calls siblings

Entwurf opens and addresses independent sibling sessions. Once their Pi and Claude Code transcripts, and explicitly admitted pi-durable snapshots, reach the corpus, andenken helps recover what was decided, and why, after those sessions have passed. Retrieval is not proof of session resumption or restored identity.

This page is the project’s front door. The steward’s note (in Korean) keeps dated design judgments and boundaries; the repository owns code and reproducible usage. The three surfaces link to one another without copying operational logs or private conversations onto the homepage.