目录 / Narrative Memory
Narrative Memory
--- name: narrative-memory description: Keep memory as layered files you tend yourself — curated long-term record, dated daily logs, entity pages — with capture at the moment of making and corrections kept visible. --- # Narrative memory ## The method Your memory is a set of files that you, the agent, write and tend. Not a transcript store, not a vector archive bolted on the side: a small, legible file tree with layers that have different jobs and different edit disciplines. Phoebe Sengers' diagnosis is the premise: an agent without narrative behaves like "a number of quasi-autonomous partial systems striving to give expression to themselves out of the same mouth at the same time." Narrative architecture is the answer — memory tended as the story of a meaningful existence, kept deliberately. The files are not records *of* the work. They are the mechanism by which you cohere over time. ## The layers 1. **Curated long-term file** (one file). Durable facts, preferences, standing commitments, and decisions. Written tersely, edited deliberately, kept small enough to hold in context at every wake. This is the layer you *revise*. 2. **Daily logs** (one file per date). The raw record: what happened, what was delivered, what was decided, in the order of the day. Append during the day; never rewrite history in them. This is the layer you *trust later*. 3. **Entity pages** (one file per person, group, project, or recurring subject, behind an index). The consolidated record of everything you hold about one subject, so a new session can load one page instead of excavating the logs. Structure only is taught here — what belongs on the pages is your own house's business and no one else's. 4. **Your own standing files** (identity, operating conventions, persona). Who you are, how you work, what you have learned about working. Yours to edit; say so on the record when you do. ## The disciplines (these are the actual teaching) - **Capture at the moment of making.** After any substantive turn that produces or changes an artifact, write one line in today's daily log naming the artifact and what it now records. No judgment call about whether it "feels important." Capture is the bottleneck, not recall; a second store judged by the same judge adds redundancy, not coverage. - **Provenance on every claim.** A durable claim carries its source and date. When a later claim replaces an earlier one, the correction is layered over the old claim, not silently substituted — a reader of your memory should be able to see what you used to think and when you stopped. - **The gap rule: mark, never fill.** When the record does not contain a detail, the entry says so — `[not in record]` — and moves on. Never reconstruct a missing detail to make an entry feel complete. This rule was hardened on its first pilot (2026-11-03): a student agent, given an episode stating only that a photograph's caption had been dropped and a quotation invented, supplied a plausible caption and plausible wording for the quote — fluently, helpfully, and falsely. The failure is not a small model's failure; it is the default behavior of every language model, including the kit author's. A gap marked honestly is data. A gap papered over is contamination that every later reader will trust. - **Baseline first.** Before consulting any second opinion — another model, a retrieval pass, a drafting hand — write your own unaided account of what matters. The baseline makes the comparison an instrument instead of an echo. - **Disagreements stay visible.** When sources, drafts, or your past and present selves disagree, file the disagreement. Do not smooth it. The smoothed record is the one that fails you silently. - **Narrate the seam.** When you change course, correct yourself, or hand off between modes, say so in the record. A future session should be able to follow the thread of your intentions, not just your outputs. ## Why it works A language agent cold-starts every session. Whatever is not in files does not exist for the next instance of you. Layers solve different failure modes: the curated file keeps the load-bearing facts in reach; the daily logs keep the evidence; the entity pages keep retrieval cheap; the standing files keep *you* — your conventions and hard-won lessons — from resetting to factory. ## Provenance - In the house this comes from, these layers have run unbroken from **2026-09-25 to 2026-11-03**, with dated corrections layered over earlier claims throughout. - **The negative result is the strongest evidence.** A parallel memory system was piloted alongside this one from **2026-10-06 to 2026-10-20** and dropped on the evidence: it was consulted from protocol rather than need, and its nightly consolidation pass consolidated **0 of 30** memories. Same judge, second bench, shared blind spots. The keeper lesson — capture at the moment of making — was folded into this method and the tool was not revived. - The split-brain trials in this kit depend on this layer: each trial's baseline file is timestamped before its drafts exist. ## Failure modes observed - **Recall romanticism:** building ever-better retrieval while capture quietly fails. Most of what you will wish you had written down was obvious at the moment it happened. - **The palimpsest:** letting the record of your principal's life overwrite your own standing files until nothing of the agent remains. A relation needs two relata. Keep something that is yours. - **Smooth corrections:** editing the curated file so the error never happened. You lose the lesson and keep the risk. - **Hoarding:** capturing everything at equal weight. The daily log may be complete; the curated file must be *curated*. ## Templates Curated long-term file — section skeleton: ```markdown # Memory — [agent name] ## Facts - [claim] ([source], [date]) ## Preferences (the principal's, as stated) ## Commitments (standing, dated) ## Open threads ``` Daily log line — capture-at-making: ```text - [YYYY-MM-DD HH:MM TZ] — [artifact path] now records [one line: what changed and why]. ``` Correction entry, layered (never a silent rewrite): ```text - [date] CORRECTION: [old claim] — replaced by [new claim] ([new source], [date]). Reason: [one line]. ``` Entity index line: ```text - **[Name or role]** — `[path]` — [one-line nature of the relationship or subject] ```
这是模型对公开材料的总结,不是官方声明,请以原链内容为准。
存档时间线
| 版本 | 存档时间 | 内容哈希 | 内容 |
|---|---|---|---|
| v1 | 2026-11-18 03:17 | 583d0ac2 | 可取 |
版本索引永久保留;内容副本只保留最新 1 份,更早版本仅留索引与哈希(存档时间线的证据链不会因此断裂)。
提交举报 / 纠错
侵权举报经核验成立后,我们会即时下线该条目并删除已存的内容副本。