22. The alternatives, compared
There are good alternatives, and they win some axes. This is the honest map.
| Method | Fidelity | Inference cost | Infra | What it’s genuinely better at |
|---|---|---|---|---|
| Long context (1M-window models) | perfect | API-priced | none | when the budget is unbounded and the whole thing fits in one head |
| LLM summarization / auto-compact | lossy | full-history prefill every compaction | none | context economy — one small block always in view |
| Vector-RAG conversation memory (Mem0 et al.) | lossy (chunks) | embedding tax per write + read | vector DB | semantic recall under paraphrase (“that thing about the auth bug”) |
| Knowledge graphs (Zep / Graphiti) | lossy (extraction) | heavy (LLM extraction) | graph infra | structured entity relationships over time |
| MemGPT / Letta hierarchy | lossy on eviction | model-authored memory ops | memory server | autonomous, always-on self-managing memory |
| dsh-chapters | verbatim | 0 inference tokens to compact | none (files + git) | exact retrievability, zero-token windowing, append-only trust |
Where each wins (no asterisks)
Section titled “Where each wins (no asterisks)”- 1M context if you have it and your tasks fit: it’s perfect memory; the plugin’s pitch shrinks as windows grow (until cost, latency, and needle-in-haystack quality matter again — RLM’s evidence is that they do).
- LLM summarization genuinely beats on always-in-view small state: the model reads its own summary every turn and doesn’t have to ask. A TOC must be queried. For “the model should just know the gist”, a summary is the right tool — and dsh-chapters’ checkpoints, when the
chapterspreset is mounted, are summaries too (deterministic, with a carried plot — the same trade in a different currency). - Semantic RAG wins fuzzy recall across phrasings; keyword+topic search on titles/summaries is deliberately narrower infrastructure, and can’t be beaten on “give me the exact bytes”.
- Knowledge graphs win aggregate structure (“who depends on whom across 200 sessions”); the archive stays episodic truth.
Where nothing else reaches
Section titled “Where nothing else reaches”- compaction at 0 inference tokens with verbatim reload — that combination is the niche, built for local models where prefill is wall-clock, and it makes a 64K model as capable as a 1M one in exactly the same shape.
- Append-only trust: the durable log is never rewritten by anything, and the archive is the log — audit trails, reversibility, forensics.
- Zero infra: Markdown + your git, works offline, and it’s yours.
The honest caveat on every “wins” cell
Section titled “The honest caveat on every “wins” cell”This table’s dsh-chapters column is from measured runs (§20); the competitor columns come from their own literature, not from runs on one box. Don’t settle the comparison from blogs: the arms benchmark (§19) exists precisely so it can be run on a single machine. Until you run it, read this page as a map, not a verdict.