SynapCores Community Edition v2.3.1-ce — DAVINCI

Community Edition · free · seamless upgrade from v2.2.0-ce

Your database now learns from outcomes, and proves every decision.

DAVINCI brings governed autonomous learning directly into the database. Agents learn which actions deliver verified results, retain useful experience, and improve recurring workflows over time. Every decision has a tamper-evident history. Your SQL data, graph, vectors, agent memory, and inference remain together under one security boundary.

This release also helps teams scale vector search beyond available RAM, reuse existing Ollama integrations, ask grounded questions of live operational data, and deploy AI workloads with fewer moving parts.

The result: less integration work, more predictable infrastructure costs, and agent workflows that are measurable, auditable, and controlled.

1. Agents that improve from verified outcomes

Outcome: Reduce repeated prompt tuning and improve recurring task execution.

Define a goal and a deterministic success check. SynapCores executes the goal, records the result, and updates an online policy based on verified outcomes. Policy promotion requires separate held-out evidence, helping prevent unproven changes from reaching active workflows.

This is online policy learning, not foundation-model fine-tuning. Learning stays within your tenant and does not send learning data to an external model vendor.

Built-in controls:

• Off by default, with opt-in per goal or persona.
• Explicit approval before write actions. Pending writes leave underlying rows unchanged.
• Per-goal spending limits.
• Durable pause and resume controls, including a kill switch that survives restarts.

Verified in the release image: Goal execution persisted a scored episode and updated the online policy. A write request returned awaiting_approval without changing the row. The kill switch persisted across a halt, and promotion without disjoint held-out evidence was rejected.

Developer entry points: CREATE GOAL, EXECUTE GOAL, SHOW AGENT GOALS|POLICY|EPISODES, PAUSE|RESUME GOAL, and AGENT_RUN(persona, task, JSON_OBJECT('learn', true, ...)).

2. Prove what your agents did

Outcome: Make autonomous operations traceable and tamper-evident.

Every recorded agent decision, reward, and policy change enters an append-only, cryptographically hash-linked ledger. DAVINCI verifies the full agent and episode history, including individual record lineage, rather than checking only the latest entry.

Review the sequence of decisions, inspect the evidence behind learning, and detect changes that break ledger integrity. This supports internal investigations, governance reviews, and audit preparation.

Verified: End-to-end chain verification passed with zero verification failures across the tested agent and episode history, including floating-point payloads.

Scope note: A hash-linked ledger provides tamper evidence. Regulatory compliance and resistance to a privileged attacker require additional controls, such as access restrictions, external anchoring, and independent retention.

3. Grow your vector corpus without growing RAM at the same rate

Outcome: Control memory costs while keeping similarity search available.

The new elastic HNSW vector store retains frequently accessed vectors in memory and pages other vectors from durable storage within a configured memory budget. Keep relational data, graph relationships, and embeddings in one engine rather than operating a separate vector service.

Verified: A corpus sized at 1.5× the memory budget completed testing across 3 tenants and 64 checkpoints, with zero out-of-memory failures. Measured SQL and vector-search latency stayed within ±6% of the prior release in the validation workload.

4. Keep your Ollama clients and existing AI workflows

Outcome: Reduce migration effort and preserve application integrations.

SynapCores now supports native Ollama API workflows tested with ollama-python, LangChain Ollama, and LlamaIndex embeddings. Redirect supported clients to SynapCores without rebuilding their integration layer.

Reliability improvements include request queuing for concurrent /api/generate calls, proper 404 responses for missing collections, safe handling of null generation options, and corrected model reservation when num_gpu: 0.

Verified client matrix: ollama 0.6.3, langchain-ollama 1.1.0, and LlamaIndex embeddings.

5. Ask questions of your operational data, with evidence

Outcome: Give teams natural-language access to live data while retaining a path to verification.

Ask a question scoped to a database. SynapCores uses the live schema and rows to produce an answer accompanied by the generated SQL and retrieved evidence. Improved column-key enforcement, aggregate handling, and case-tolerant text predicates reduce common query failures.

For RAG and agent answers, selected-source grounding is stronger, source references are readable, and tool failures no longer surface as if they were answers. File citations now identify a meaningful path, such as filesystem:orbit_support/troubleshooting-error-codes.md#0, instead of an opaque identifier.

General chat no longer invokes database or RAG tools by default. Tool use is opt-in per turn. The chat service also respects [query.ai_service].model across configuration tiers.

Why this matters: Users see where answers came from. Engineers have the SQL and source material needed to check them. Grounding reduces unsupported answers, but does not eliminate model errors.

6. Give agents more accurate, up-to-date memory

Outcome: Improve semantic recall and prevent outdated facts from taking precedence.

CREATE MEMORY now matches vector dimensions to the selected embedding model, including 768-dimensional models such as nomic-embed-text, instead of assuming 384 dimensions. Explicit model and dimension settings remain available.

Memory supersession now respects the chronology of updates. When a current fact replaces an older fact, the system preserves the newer state rather than incorrectly prioritizing stale history.

Verified: 768-dimensional memory creation and recall, a Miami-to-Denver update, rejection of a historical claim, and successful reopening of existing 384-dimensional stores.

7. Start with a better default for RAG and agent tasks

Outcome: Improve the new-install experience without requiring model selection.

New deployments default to qwen2.5:7b, a general instruction-following model selected for RAG and agent workflows rather than the prior coding-focused default. Existing installations with an explicitly configured model retain their selection.

Verified: The default model downloaded, loaded, and generated a response on a CPU-only host without AVX-512 support.

Model-quality improvements are a release objective, not a published comparative benchmark.

8. Spend less time configuring and repairing deployments

Outcome: Simplify setup and lower upgrade risk.

The engine detects available GPU hardware without manual layer configuration and continues to support CPU-only hosts. Installer improvements reconcile stale service drop-ins, support NVIDIA device nodes, and preserve customer-owned configuration during uninstall.

Existing memory and vector stores remain readable, with no on-disk format change required for this upgrade. Corrected ML recipes 002, 004, 005, 006, and 007 run end to end on v2.3.1.

Install or upgrade

docker pull synapcores/community:v2.3.1-ce

Or use the installation script:

curl -fsSL https://get.synapcores.com/install-ce.sh | sh

Release binaries are available for Linux x86_64, Linux aarch64, and macOS Apple Silicon. Review your deployment configuration and back up production data before upgrading.

Community Edition. No separate AI stack required.

All capabilities described above ship in SynapCores Community Edition. Autonomous learning stays off by default until explicitly enabled.

One database for SQL, graph, vectors, memory, inference, and governed agents. Learn from verified outcomes. Keep control of every action. Prove the history.