SerenLoci

The left brain. A keyed facts-and-logic store for the Seren constellation - addressable, deterministic, exactly one live value per key.

Where SerenMemory (the right brain) holds fuzzy, episodic memory - "we ground on the embedder migration for a week, it was a slog" - Loci holds facts:

{ project, key, value, why }

camelCase is life. braces on a new line in posh. GGML_CUDA_NO_VMM=ON must be set at compile time - because the env var isn't honored at runtime on Jetson. A locus has an address. You go to it and get the thing - you don't grope around for something that rhymes.

Why it's shaped like this

One live value per key, enforced by the database. Set a new value for a key and the old one is superseded - kept as history, pointed at by the new row, never blended. A vibed-together fact is worse than no fact, so the rule is strict, and it's not enforced by hope: a PARTIAL UNIQUE INDEX makes sqlite physically refuse a second live row per (project, key).

Two tiers. A reserved project of * is the fundamentals tier - cross-project truths. A concrete project name (seren-memory) is the per-project tier - the targeted override. Same split as nano/xavier/dgx prebuilts: the platform-wide truth and the per-board variant.

The why is the point. A logic store without rationale is just a dictionary. The value tells you what; the why is what stops you re-learning it the painful way - and it's searchable, so "that CUDA thing" finds a fact whose reason mentions CUDA even when the key doesn't.

The floor is free

Three access rungs, cheapest first:

1. Exact - get_fact(project, key) returns the live value, deterministically. No embedding, no ranking. You know the address, you get the thing.
2. Lexical - FTS5 full-text over (key, value, why) of the live rows. The "I sort of remember the words" path.
3. Vector (additive) - a sqlite-vec index for the "this smells like that CUDA thing" associative jump. Built only when you name an embedder.

Rungs 1 and 2 need nothing but sqlite (stdlib) and the web stack - no torch, no GPU, the 4GB-laptop floor. The vector finder is the ceiling, opted into by install, never required. Where does it sit? How much does it need? - the floor needs almost nothing.

Install

pip install seren-loci                 # the floor: exact + FTS5 lexical
pip install seren-loci[vector]         # + the sqlite-vec associative finder (pulls torch)
pip install seren-loci[mcp]            # + the MCP surface (model reaches facts directly)
pip install seren-loci[corp]           # + OS-trust-store TLS for corp-proxied boxes

Extras stack: pip install seren-loci[vector,mcp].

Run

seren-loci                             # or: python -m seren_loci
seren-loci --config ./seren-loci.yaml

Listens on 7422 by default (neighbor convention: memory 7420, margin 7421, loci 7422).

# set a fundamental truth
curl -X POST localhost:7422/fact -H 'content-type: application/json' \
  -d '{"key":"posh.brace_style","value":"curly brackets on a new line","why":"readability"}'

# get it back, deterministically
curl 'localhost:7422/fact?key=posh.brace_style'

# discovery search (exact-key first, then the finder)
curl -X POST localhost:7422/search -H 'content-type: application/json' \
  -d '{"query":"cuda runtime"}'

API

Every search hit carries a normalized 0-1 score (exact->1.0, hybrid->RRF fused into 0..1, lexical->bm25 mapped above 0). That's the common currency SerenCorpusCallosum uses to merge left-brain and right-brain results on one axis instead of comparing cosines to key-hits.

The hybrid finder (additive - the ceiling, not the floor)

When an embedding model is configured, rung 3 is no longer a pure vector KNN. Instead a _HybridFinder runs both lexical (FTS5) and semantic (vector) search in parallel, then fuses results with Reciprocal Rank Fusion (RRF):

rrf_score(fact) = Σ 1/(k + rank_in_ranker)      # k = 60
                 ranker ∈ {FTS5, vector}

This gives better precision than either ranker alone - FTS5 catches exact/keyword matches with high precision, vector catches semantic matches with high recall, and RRF boosts documents that rank highly in both lists. The old pure-vector path (_VectorFinder) is replaced by _HybridFinder whenever an embedder is present; the floor (no embedder) still uses FTS5 only.

Scope-aware query embedding. When a single project scope is provided (and it isn't the * fundamentals tier), the vector query is augmented by prepending the project name before the raw query string. For example, searching "rate limiting" under project seren-memory embeds "seren-memory: rate limiting" instead - nudging the query vector toward the memory-domain cluster so it disambiguates facts that share wording across projects.

Tests

Run with pytest (sqlite-vec-gated tests need the [vector] install or a PYTHONPATH into the vector venv's site-packages):

pytest tests/                                    # floor tests (no vector needed)
PYTHONPATH=/path/to/vector-venv/lib/python3.12/site-packages pytest tests/  # all tests

Where it sits in the constellation

• SerenMemory - the right brain. Fuzzy, consolidated, episodic. General-purpose AI memory protocol.
• SerenLoci - this. The left brain. Keyed facts, deterministic, strict-supersede.
• SerenCorpusCallosum - fans a query across both hemispheres and merges on the shared score currency.

Build for the floor, not the ceiling. The Nano is the floor, not the cap. GPL-3.0-or-later. Rip it and win.