Optional Hindsight implementation

This is the optional memory and synthesis configuration entrance in the reference and method overview. Your own practice starts from Empty, which includes this configuration guide. Hindsight assists source retention, Observation integration, and maintained Mental Models/Knowledge Pages for six questions. The model configures the service, checks its output, and adopts the useful results.

Configure the selected native implementation

Resolve the installed/supported Hindsight version, deployment form, service URL, authentication, private bank, storage, inference and embedding providers, and phase-specific model settings for the actual host. Use the upstream project for its current installation route; these choices vary across environments. The native interfaces below were inspected in version 1.1.0, not assumed valid for every installation. No service or credentials are provisioned by installing the Yinyang Skill.

1. Choose the actual native service or hosted deployment and a private source scope. Configure its persistent database/storage, inference and embeddings; keep credentials with their existing secret owner, outside the Skill and ordinary Git files. Establish actual access and provider/data boundaries.
2. Read the effective bank configuration, then set retain_mission, observations_mission and reflect_mission for the shared purposes below. Set the phase-specific provider/model/budgets and source retention appropriate to the actual data and cost; do not copy another person's runtime settings.
3. Define the six complete questions from all six guides and actual grounds, configure native scope and full/delta refresh, and connect the intended caller through its supported SDK/API or MCP route. Questions remain related, not six isolated corpora. Record the actual binding and reading locations.
4. Check service health and bank/config retrieval, then test native retention, extraction, consolidation, reflection, page refresh and Markdown export with a disposable nonpersonal record. Inspect terminal operations and read actual content back; endpoint acceptance is not the complete result.
5. Process authorized real sources through the preserved revision dependency, inspect the six-account whole and project checked bodies for the actual reader. Verify the real consumer and necessary environment effects. A route failure is an explicit native-integration gap; the simple route remains a supported independent choice, not evidence that the native operation worked.

Prefer projection to the familiar six references/situated/{orientation,person,counterpart,relationship,environment,revision}.md paths when that fits the consumer. If another native layout is useful, keep references/situated/index.md as the explicit private service/bank, six question IDs, source-scope and complete-body binding. It must expose all six reading roles, with the actual situated revision method at revision.md or its explicit complete binding. A binding selects bodies and cannot replace them. Do not keep two competing authoritative sets. Persist intended private source/export content through its actual private Git/GitHub owner; native database backup and history retain their own runtime owner. Hindsight is not itself GitHub publication.

Prepare sources and processing instructions

The selected native route preserves the revision dependency: read adopted philosophy, retrieve relevant new attributable sources, refresh and correct Personal understanding with its conflicts and uncertainty, then let Hindsight synthesize the six, check their coherence and realize necessary Skills/environment changes with actual readback. Native operations implement these responsibilities; access to original sources alone does not replace the Personal-correction step. No complete manual rewrite or mandatory chain of generated summaries is required. All products retain routes to original grounds.

Philosophy supplies arguments, distinctions, and chosen commitments; Personal supplies attributed reports and evidence; environment and work/history records supply conditions, events and actual results. Preserve author, date, source locator, context and uncertainty. A chosen philosophy is distinct from a hypothesis about a person. Original sources retain their evidence and authorship owners. Personal also remains a human-readable collection of substantive topic articles in its existing four subject groups. Hindsight can maintain these as Knowledge Pages with detailed topic questions and project their Markdown bodies into the existing private reading paths. Keep the readable index, articles and source links usable; observations remain backstage rather than replacing articles with raw database lists. The Personal guide owns that reading purpose and organization.

Migrate affected derived articles when actual revision needs it, preserving unique evidence and current readability. No wholesale Personal deletion or rewrite, complete duplicate summary, or mandatory manual-first stage follows. The person and authorized agents may still edit and organize sources and articles directly. Personal topic pages and collaboration models derive their different content from relevant original grounds; do not require collaboration models to summarize generated Personal pages. Use native scope and sibling-model exclusion deliberately to avoid circular or self-corroborating synthesis.

In Hindsight's native categories, World and Experience are parallel raw fact categories: World includes attributed human reports, preferences, rules, and external events; Experience records actions or observations actually made by the assistant. A person's reported experience usually enters as attributed World, not assistant autobiography. Observations are revisable consolidated facets with source links. Mental Models are separately maintained documents generated by a query. None of these labels certifies truth or gives authority.

Use native bank retain_mission, observations_mission, and reflect_mission for shared extraction, consolidation, and synthesis purposes. Retain relevant relations, conditions, alternatives and correction history rather than only traits. Keep phase guidance coherent: a relation omitted during extraction may be unavailable to synthesis. Tags, tag groups and fact_types scope retrieval; they do not encode all philosophical relations or an instruction hierarchy. Native directives are explicit instructions for reflect/refresh; promote only actually adopted requirements deliberately, never every inferred preference. They do not automatically become Codex instructions or authorize host changes.

Write model questions as detailed content specifications

Keep six explicit native model questions: orientation, person, counterpart, relationship, environment and revision. Their full functional scope belongs to the six-question method and the six complete formation guides. Read those guides together when defining or materially revising each source_query, so its scope preserves its own subject and the relations it has to the other five. A heading or a shortened role list does not supply the guides' full content.

A question can specify purpose and consumer, subject scope, important relations, priority among concerns, organization, evidence distinctions, uncertainties, counterexamples and how the account should guide work. Its prose organization can vary with the material. Do not merge person, counterpart and relationship into one undifferentiated question or treat environment as only an inventory. The revision question must form the person's actual method for reconsidering purposes, material, interpretations, connected changes and verified use; native API procedures alone do not supply it. The six philosophical dimensions remain questions of understanding, not a database storage schema.

Specify the shared adopted premises and cross-question relations in the relevant questions while preserving each question's distinct contribution. For example, a chosen aim of voluntary participation must remain intelligible in the person account, wanted counterpart contribution, allocation of joint work, enabling environment and means of correction. This calls for connected reasoning, not six copies of one paragraph.

Check coherence across the six questions

Read all six complete accounts together after formation and a consequential revision. Check shared premises, concepts and referents, conditions and time, responsibilities and practical consequences against sources and actual use. A common bank, shared missions, retrieval from the same sources, or a successful refresh for every page does not guarantee that these relations are coherent. Preserve real disagreement or uncertainty instead of smoothing it away.

Sibling accounts may be read as current interpretations to coordinate. They are not independent evidence for one another's factual claims. Inspect original grounds where a claim matters, and distinguish using a sibling's stated view as context from counting it as corroboration. Configure native model inclusion or exclusion deliberately to avoid echoing a stale view; do not turn exclusion into a requirement that questions be understood or revised in isolation.

If the whole exposes a mismatch, repair the affected original source, mission, query, retrieval scope or environmental realization, refresh the dependent questions, and read the six together again. A local correction need not regenerate unaffected bodies, but its consequences must be checked across the whole. Keep important unresolved differences explicit. The situated revision method carries this judgment and application; no mandatory seventh summary document is needed.

Use the native interfaces

The following interfaces were inspected in Hindsight 1.1.0's installed Python client and API implementation. Verify them against the installed version when configuring another host. The upstream project owns installation and evolving provider/service instructions.

from hindsight_client import Hindsight

client = Hindsight(base_url="http://localhost:8888")  # actual configured URL
bank = "private-example"
client.retain(
    bank_id=bank,
    content="An attributed, authorized source passage with its context.",
    document_id="source-owner-record-id",
    context="Source identity, date, evidence role and applicable scope",
    tags=["philosophy", "personal"],
)
created = client.create_mental_model(
    bank_id=bank,
    name="The actual model counterpart",
    source_query="""Develop the counterpart question from the six complete
    formation guides and relevant original grounds. Explain actual capabilities
    and conditions, wanted contribution, independent judgment and fallibility.
    Relate them coherently to adopted direction, the actual person, joint work,
    environment and situated revision. Distinguish observed effects, intended
    arrangements and hypotheses; preserve conditions and unknowns.""",
    max_tokens=6000,
    trigger={"mode": "full", "fact_types": ["world", "experience", "observation"],
             "exclude_mental_models": True, "refresh_after_consolidation": True},
)
# Creation schedules work: inspect its operation_id and terminal result,
# then read the resulting ID through client.get_mental_model(bank, model_id).
# Specify the other five full questions and check all six accounts together.

The example illustrates fields, not a command to run against arbitrary personal data or a universal model/provider/token choice. The SDK also exposes create_knowledge_page(bank_id, name, source_query, ..., trigger=...), refresh_mental_model(bank_id, mental_model_id, mode="full"), reflect(bank_id, query, ...), and export_knowledge_base(bank_id). Use a... async counterparts in async callers; generated API clients are also available through client.banks, client.mental_models, and client.knowledge_base.

Native REST routes under /v1/default/banks/{bank_id} include:

Read current config before patching; a mission update replaces its value. Observe operation completion and errors rather than treating an HTTP acceptance as generated content. Inspect source/fact extraction, consolidation backlog, resulting observation provenance and model refresh independently. Full refresh is useful after changing the question or source interpretation; select delta and automatic refresh when their actual behavior fits continuing maintenance. Native Knowledge Pages default to delta refresh from observations, excluding other mental models. Override that scope deliberately when raw-source evidence is needed. Excluding sibling models can prevent an old synthesis being treated as fresh independent support. Export the native Markdown bundle when a file consumer needs it; no second custom synthesis engine is required.

Check the formed understanding and its use

Read the complete generated bodies alongside representative and consequential sources. Check that the chosen philosophy, personal reports, practical relations, conflicts and uncertainty survived, and that quoted history has not become the current assistant's experience. Inspect gaps through recall/source retrieval; repair source context, missions, scope or model questions, rerun the native operation and read the changed result. A smaller output or more observations does not prove semantic preservation. A source citation is not independent corroboration, and fluent synthesis does not establish a causal explanation.

After checking the complete six-question whole, update the private index's actual binding and required startup reading. Migrate physical consumer paths through their owners only after the successor preserves the required functions and relations. Existing actual bodies can remain the current reading source until that replacement is checked; retirement is not automatic on generation. Check the real calling conversation can retrieve the selected complete bodies. Realize environment changes through their existing owners and verify actual use. The execution owner retains purpose, judgment, authority and completion; Hindsight supplies formation and revision machinery. Source delivery, native fusion, consumer availability, useful cooperation and human development establish different claims and need evidence at their respective scales.