Skip to content

Roadmap: SQLite vector quantization at scale; drop competitor numbers #260

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
samrusani merged 1 commit into from
Jul 6, 2026
Merged

Roadmap: SQLite vector quantization at scale; drop competitor numbers #260

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 3 additions & 2 deletions ROADMAP.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
# Roadmap

## Baseline (Not Roadmap Work)
- **LongMemEval_s: 64.6%** published with full methodology and per-question evidence in [docs/benchmarks/longmemeval/](docs/benchmarks/longmemeval/README.md) (Zep reports 63.8%; Mem0 ~49%).
- **LongMemEval_s: 64.6%** published with full methodology and per-question evidence in [docs/benchmarks/longmemeval/](docs/benchmarks/longmemeval/README.md). We don't quote other tools' numbers — published runs differ in response models, pipelines, and dataset revisions; our per-question receipts are committed so the run can be judged directly.
- `v0.9.0`: released. Complete Memory Operations Protocol (all ten verbs, including true `redact`), Context API v2 (per-section budgets, packing strategies, depth tiers), full export/import round-trip, temporal graph memory + entity resolution with the `entity_resolution` and `graph_hop_retrieval` eval suites, and the published scale envelope.
- `v0.8.0`: released. The memory-frontier waves — budgeted context packs, typed retrieval, staleness v1, agentic write protocol on the core MCP surface (11 tools), memory-quality eval suites, real scopes with key binding, merging consolidation, temporal slice, LongMemEval harness — plus Alice's first published benchmark result.
- `v0.7.0`: released. Zero-infrastructure SQLite on-ramp — `uvx alice-memory mcp` serves the core tools with no Docker or Postgres; published to PyPI via Trusted Publishing.
Expand All @@ -14,7 +14,8 @@ In rough priority order:
1. **Multi-session synthesis** — the weakest LongMemEval category (45.1%): retrieval breadth across sessions plus consolidation-driven aggregation; benchmark-driven iteration with the per-type breakdown as the scoreboard. The breadth ablation (49.2% at 2× context) motivates the planned query-shape-aware aggregation mode.
2. **Dogfood daily** — run the stack against real agents with an embedding endpoint; calibrate the paraphrase target of the `retrieval_quality` benchmark; generate the usage telemetry that future ranking/policy improvements need.
3. **Reference integrations** — deeper examples for popular agent frameworks on the core tool surface.
4. **Hosted offering exploration** — the RLS posture and auth work make this plausible; still exploratory.
4. **SQLite vector search at scale** — push the local-first on-ramp's embedding search past its documented ~20–30k comfort zone (brute-force cosine hits ~2.3s at 100k memories in the [scale envelope](docs/benchmarks/scale/README.md)): quantize stored embeddings (int8/binary) and rescore the top candidates at full precision, with the recall deltas published alongside the latency wins. Proven techniques, no new dependencies; recent results like Google's TurboQuant line (ICLR 2026) show how much headroom few-bit vector quantization has, and serve as the reference point for how far to take it.
5. **Hosted offering exploration** — the RLS posture and auth work make this plausible; still exploratory.

## Explicit Non-Goals For Now
- Hosted service / SLA commitments.
Expand Down
Loading